supplementary information - nature...112 due to the orbital geometry and short separation between...
TRANSCRIPT
![Page 1: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/1.jpg)
SUPPLEMENTARY INFORMATIONDOI: 10.1038/NGEO2513
NATURE GEOSCIENCE | www.nature.com/naturegeoscience 1
Supplementary Information1
Strong glacier mass loss in the Tien Shan2
over the past 50 years3
Daniel Farinotti1,2, Laurent Longuevergne3, Geir Moholdt4, Doris Duethmann1,4
Thomas Molg5, Tobias Bolch6,7, Sergiy Vorogushyn1, and Andreas Guntner15
1 GFZ German Research Centre for Geosciences, Section 5.4 -Hydrology, Potsdam, Germany6
2 Swiss Federal Institute for Forest, Snow and Landscape Research WSL, Birmensdorf, Switzerland7
3 National Center for Scientific Research CNRS, UMR 6118 Geosciences, University of Rennes, France8
4 Norwegian Polar Institute, Fram Centre, Tromsø, Norway9
5 Climate System Research Group, Institute of Geography, Friedrich-Alexander-University Erlangen-10
5 Nurnberg (FAU), Erlangen, Germany11
6 University of Zurich, Department of Geography, Zurich, Switzerland12
7 Technische Universitat Dresden, Institute for Cartography, Dresden, Germany13
Substantial glacier mass loss in the Tien Shan over the past 50 years
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 2: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/2.jpg)
A Estimates based on GRACE data14
Launched in March 2002, the Gravity Recovery And Climate Experiment (GRACE) [1] has15
revolutionized the way large mass changes can be detected on Earth. By monitoring the temporal16
variations of Earth’s gravity field with an “unprecedented temporal and spatial resolution” [2],17
GRACE has provided new insights in mass redistribution processes of the atmosphere, the18
oceans, terrestrial water, and the cryosphere [for a review see 3].19
Consisting of two twin satellites flying in a polar orbit at about 450 km altitude and about20
200 km apart, GRACE infers on Earth’s gravity variations by constantly monitoring the distance21
between the two satellites at the micrometer level [e.g. 4]. Two types of products have been22
developed from GRACE range-rate data: The first translates satellite range-rate data directly23
into a set of localized surface mass concentrations [so-called ”mascons”, e.g. 5]; the second is a24
global spherical harmonic (SH) expansion of the gravity field, where a set of Stokes coefficients25
[e.g. 6] is the standard GRACE Level 2 product. Note that mascon- and SH-derived mass changes26
are equivalent [e.g. 7, 8]. In the present study, the SH formulation is used. A description on27
how surface mass changes can be derived from SH coefficients is given in [9]. Love numbers by28
[10] are used.29
A.1 General processing flow and challenges30
When recovering surface mass variations from GRACE products, two major difficulties have to31
be dealt with. The first is linked to the limited spectral content (i.e. the sensitivity to large32
spatial scales) of GRACE when focusing on a space-limited areas [11]. The second is given33
by the fact that GRACE (and more generally, gravity) provides information about vertically34
integrated mass changes only, which makes a separation of individual sources challenging.35
The first point leads to so-called leakage effects [12], i.e. to a loss in signal amplitude when36
concentrating GRACE on a region of interest, and a partial compensation from mass changes37
outside that region. Several methods have been developed to overcome this issue, including38
the use of a-priori information on the spatial distribution of the expected mass changes [e.g.39
13, 14]. The second point is inherently difficult, and is generally tackled by using models40
to discern between individual sources. For the region of interest in this study, the processes41
that potentially contribute to gravity variations can be subdivided into two categories, i.e. (1)42
processes related to near-surface mass transport, such as water storage in lakes, the unsaturated43
zone, aquifer systems, glaciers, the seasonal snow cover, and erosion, and (2) internal processes,44
including glacial isostatic adjustment since the Last Glacial Maximum and the Little Ice Age,45
and vertical crustal movements related to tectonic processes.46
Previous studies focusing on GRACE-derived glacier changes in High Mountain Asia showed47
that the largest uncertainties stem from the hydrological contribution [15, 16]. Recently, [17]48
showed that forward modelling (i.e. application of a spatial filter to the modelled hydrological49
contribution in order to mimic the large-scale sensitivity of the GRACE signal) can be used50
for reducing that uncertainty. Given that the processing method used for the derivation of51
the GRACE data is known, the required mathematical process is straightforward. Forward52
modelling the impact of all known contributions to the GRACE signal was previously shown53
to be the most suited method for extracting a specific storage compartment [e.g. 18, 19]. This54
general strategy is adopted here. In principle, the impact of all known mass-change contributions55
derived from models and/or independent estimates (Sections A.3 and A.4) are spatially filtered to56
match the GRACE resolution, subtracted from the total mass change derived from the GRACE57
data (Section A.2), and the residuals interpreted as glacier mass changes (Figure A.1). The so58
obtained glacier mass changes, which still refer to the GRACE resolution, are re-focussed on the59
region of interest by using a mascon adjustment approach (see Section A.5). Mass change rates60
are then obtained by computing time-series trends through a 4th order low-pass Butterworth61
filter [e.g. 20] in order to remove seasonal variations. In order to account for a wide range of62
2
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 3: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/3.jpg)
possible error sources, the workflow is based on an ensemble-like approach that accounts for63
uncertainties in the (1) GRACE-derived gravity change products, (2) applied spatial filtering64
and processing strategy, (3) mass-change contributions other than glaciers subtracted from the65
total signal, and (4) methodology for re-focusing the residual mass changes.66
A.2 GRACE data67
Different sets of Stoke coefficients (called ”GRACE solutions”) can be retrieved from various68
research groups. Differences reflect varying computational strategies and are related to (i) the69
use of different background models to compute and remove the contribution of atmospheric and70
oceanic mass changes, (ii) the inversion method used to estimate the Stokes coefficients, and71
(iii) the spectral content. Although a recent global analysis suggests that long-term trends are72
similarly captured by different GRACE solutions [21], significant local differences have been73
reported to occur [e.g. 22, 23].74
In this work, a suite of 12 GRACE solutions is used in order to quantify the impact of various75
processing strategies. The different solutions can be summarized into three main categories76
including (1) unconstrained solutions, (2) stabilized solutions, and (3) regularized solutions.77
Unconstrained solutions are provided by the Center of Space Research (CSR), the German Re-78
search Centre for Geosciences (GFZ), and the Jet Propulsion Laboratory (JPL). For the CSR79
and JPL solutions, product Release 05 (RL05) is used, whilst RL05a is used for GFZ. Data80
are retrieved from ftp://podaac-ftp.jpl.nasa.gov/allData/grace/ and refer to the period81
2003-2013. Compared to RL04, ameliorations in both range-rate data processing and background82
models have resulted in lower noise level in RL05 solutions, leading to a significant reduction83
of residuals over the oceans [24] and an improved effective resolution. Unconstrained solutions84
generally show prominent North-South striping related to the amplification of uncertainties in85
the background models, and the increasing noise with increasing degree [e.g. 25]. For improving86
the signal-to-noise ratio, specific post-processing filters have to be applied. Here, three types of87
common filters are explored: (1) Gaussian smoother [e.g. 26, 9], which is the most widely used88
isotropic filter, and is generally applied in combination with a decorrelation filter, (2) decorre-89
lation filter [27, 28], which is a data-adaptive polynomial filter, and (3) DDK-4 decorrelation90
filter [29, 30], which is an anisotropic filter that uses synthetic GRACE orbital geometry and91
modelled signal covariance information to smooth and decorrelate the SH coefficients. For a92
more detailed discussion of different filtering methods, and the implications for hydrological93
applications in particular, refer to e.g. [31] or [14].94
Stabilized solutions covering the period 2003-2012 are provided by the Space Geodesy Research95
Group (GRGS) and are retrieved from http://grgs.obs-mip.fr/grace. The solutions are96
designed to avoid the striping present in the unconstrained solution. This is achieved by con-97
straining degrees 31 and above towards the mean gravity field during the inversion process98
[32, 33]. Here, RL2 and RL3 are used. RL2 is used without further filtering, whilst RL3 is99
truncated at degree 60 since first analyses for the area of interest indicate a very low signal to100
noise ratio for higher degrees (not shown).101
Regularized solutions covering the period 2003-2012 [25] mitigate the uncertainty introduced by102
the ill-conditioned inversion problem by using a Tikhonov regularization. This method has shown103
to be effective and to have limited signal attenuation. Compared to unconstrained solutions,104
regularized solutions yield an equivalent representation of the original range-rate data, but allow105
the use of SH coefficients of higher degree and order (up to degree 120) with no further filtering.106
Further differences in the GRACE data can arise from the potential substitution of low degree107
terms (degrees 1 and 2). Degree-1 coefficients, which describe the relative motion between108
the Earths centre of mass and a crust-fixed terrestrial reference frame, cannot be measured by109
GRACE and are usually either ignored or substituted from time series of geocentre motion.110
Degree-2 coefficients, associated with the Earths oblateness, are affected by large uncertainties111
3
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 4: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/4.jpg)
due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the112
impact of considering/ignoring the substitution of these low degrees is quantified by using degree-113
1 coefficients by [35], and degree-2 coefficients by [36]. The analyses performed for the region of114
interest indicate that for the recovered glacier-mass change, the combined effect of substituting115
or not substituting the two coefficients is below 0.1 Gt a−1. Note that GRGS solutions use a116
specific methodology for the degree-2 coefficients [33].117
The main characteristics of the various GRACE solutions are synthesized in Table A.1. Time118
series of total mass changes for the region of interest are presented in Figure A.2a. The spatial119
distribution resulting from the use of various filtering methods is presented in Figure A.3. To our120
knowledge, neither GRGS RL3 nor the regularized solutions by [25] have been applied previously.121
A.3 Subtraction of water storage variations122
When aiming at recovering glacier-mass changes, water storage variations that need to be sub-123
tracted from the GRACE signal include variations in (1) surface water storage (SWS), (2) soil124
moisture storage (SMS), (3) groundwater storage (GWS), and (3) the seasonal snow pack (SSP).125
The combination of the three components is referred to as total water storage (TWS).126
Variations in SWS, i.e. water storage variations in rivers, lakes, and wetlands, have previously127
been ignored in GRACE-based assessments of glacier mass changes, although SWS contribu-128
tions from lakes have been shown to explain large parts of the observed mass change over the129
Tibetan Plateau for example [37, 17]. Here, the impact of SWS is taken into account following130
the methods by [38]. Outlines and corresponding surface area of major lakes and reservoirs in131
the region of interest are extracted from the Global Lakes and Wetlands Database (GLWD-132
1 and 2) [39]. Changes in lake level and/or volume are either obtained from the satellite-133
altimetry based LEGOS Hydroweb database [40], or from direct observations reported from the134
Scientific-Information Center of the Interstate Coordination Water Commission of the Central135
Asia (www.sic.icwc-aral.uz). When information about changes refers to lake-level only, vol-136
ume changes are computed using a constant-area hypothesis. For some smaller lakes without137
any available observation data, volume variations are estimated from the surface area and the138
level fluctuations of the closest lake in the same river basin. Since observations from LEGOS139
Hydroweb are available until February 2010 only, SWS contributions after that date are assumed140
to be constant. Note that this has no impact for the estimates of the 2003-2009 period, which141
are used for cross validation with other approaches (cf. Sections B and C). The time series of142
the resulting SWS contribution is shown in Figure A.2b. The characteristics of the main lakes143
and reservoirs in the vicinity (< 500 km) of the study area are given in Table A.3. Lakes that144
are included in the assessment and the contribution to TWS deriving therefrom are displayed145
in Figure A.4. Note that lake storage has generally decreased over the period 2003-2009, which146
is in contrast to what was observed for the Tibetan Plateau [e.g. 41, 42].147
The contribution of TWS has been highlighted as the major source of uncertainty when recov-148
ering glacier mass-changes from GRACE data [e.g. 15, 16, 17], and GWS and SPP are known149
to strongly influence the water cycle in the study area [e.g. 43]. Yet, the actual uncertainty in150
modelling the individual compartments has not been fully determined, as authors have generally151
been considering a limited number of models, models using the same forcing data, or models152
that do not include long-term storage compartments (GWS in particular). Here, an ensemble of153
10 land surface models (LSMs) is used to better quantify errors arising from forcing data, model154
structure, and model spatial resolution. The storage compartments (i.e. SMS, GWS or SSP)155
included in the individual models, as well as the spatial and temporal resolution of the output,156
and the data used for model forcing are summarized in Table A.2. Note that GWS is included157
in a few models only, whilst some models explicitly take into account anthropogenic water use158
as well. For a detailed description of the individual models, refer to the references in the last159
column of the same Table.160
4
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 5: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/5.jpg)
Time series of TWS variations as calculated by the individual models are presented in Fig-161
ure A.2c. The differences between individual models can be substantial. A 2-month phase-lag162
can be observed between models m05 (NOAH 2.7.1 at 0.25◦ resolution) and m02 (CLM v4.0)163
for example, whilst storage changes between individual models of the GLDAS suite (models164
m01, m04, m05, and m05a) can differ in amplitude by factor of two and more (Figure A.2b).165
Long-term water storage is generally negative, with trends for the period 2003-2009 between166
−0.3 and −8.6 Gt a−1 (average −3.7± 5.1 Gt a−1; see also the leftmost boxplot in Fig. A.6).167
In order to subtract TWS change from the GRACE signal, LSM outputs are converted in the168
equivalent SH representation, and then processed the same way as the corresponding GRACE169
solution (e.g. truncation at a given maximum degree, application of a given filter; see Table A.1).170
For the SH conversion (which uses global LSM data), Antarctica and Greenland are blanked for171
LSMs including these regions. Note that since LSM outputs do not contain the typical North-172
South stripes present in the GRACE solutions, no destriping filter is applied in order to prevent173
a potential removal of geophysical signals oriented in that particular direction [14]. The choice174
is additionally motivated by the data-adaptive nature of the destriping filters that makes the175
actual impact of the filtering procedure difficult to determine. The resulting spatial distribution176
of TWS, filtered the same way as the GRACE solutions, is shown in Figure A.5.177
Filtering the LSM outputs reduces the spread in the region-wide trends in total water storage178
(Figure A.6). This is in line with the findings by [17], and indicates that LSM uncertainty is scale-179
dependent. In particular, large-scale water storage variations seem to be better described by180
LSMs than variations at smaller scales. The strongest reduction in the spread between regional181
trends in total water storage is obtained by the strongest filtering method (truncation at degree182
60, Gaussian smoother with 300 km radius; option T60 G300 in Figure A.6). Despite this183
reduction, uncertainty emerging from the spread of individual LSM output remains the major184
contribution in the total error budget for the recovered glacier mass changes (see Section A.6,185
and Table A.4).186
A.4 Subtraction of further mass changes187
Processes other than water storage variations that affect total mass changes in the region of188
interest include Glacial Isostatic Adjustment (GIA), tectonic processes, and erosion [e.g. 15, 17].189
GIA is the ongoing viscoelastic relaxation of the Earth in response to the presence of large190
ice masses in the past [e.g. 44]. GIA contributions are generally associated with two dis-191
tinct events [e.g. 15], i.e. (1) the Last Glacial Maximum (LGM), with its deglaciation ter-192
minating about 19 000 years before present [e.g. 45], and (2) the more recent Little Ice Age193
(LIA), which had its coldest conditions between 1570 and 1730 [e.g. 46]. Modelling these194
effects require constraining both ice history and viscoelastic properties of the Earth. Two195
models and corresponding ice histories are considered here: ICE-5G [44, 47], retrieved from196
http://gracetellus.jpl.nasa.gov/data/pgr/, and RSES [48] (H. Steffen, pers. comm.,197
January 2014). Data are retrieved as SH coefficients and filtered with the same options as198
the GRACE solutions. Both the ICE-5G model, which does not show any significant glacieriza-199
tion in the Tibetan Plateau during the LGM, and the RSES model, which assumes the presence200
of small valley glaciers, suggest a negligible present-day GIA contribution for the Tien Shan201
(not shown). According to the two models, the effect on the recovered present-day glacier mass-202
change rate for the Tien Shan is below 0.05 Gt a−1. The contribution of the LIA have previously203
been estimated to be negligible by [15]. The presence of a larger glacierization during the LGM204
than what assumed in the ICE-5G and RSES glacial histories is controversial [e.g. 49, 50, 51, 52].205
[15] assessed that the present-day contribution associated to a potential Tibetan ice cap during206
the LGM can be up to 1±1 Gt a−1. Here, the contribution stemming from the ice-sheet hypoth-207
esis is not subtracted from the total GRACE signal, but the additionally introduced uncertainty208
is accounted for in the total error budget (Section A.6).209
5
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 6: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/6.jpg)
In general, the impact of tectonic uplift and erosion processes are partially compensating each210
other [e.g. 53]. Vertical uplift in the Tien Shan is difficult to constrain because of its limited211
amplitude [54], and is generally lower than for the Tibetan Plateau [e.g. 55]. The analysis of212
Bouguer gravity anomalies over the wider Tien Shan area, however, suggest that local isostatic213
compensation is nearly complete in the region [e.g. 56]. The net contribution of tectonic uplift214
and erosion is, thus, assumed to be negligible.215
A.5 Mascon adjustment216
The mass changes obtained after subtraction of the non-glacier contributions from the total217
GRACE signal (Figure A.9) still refer to the filtered GRACE resolution. For re-focussing the218
signal to the actual glacier positions, a mascon adjustment approach is used. Mascons are surface219
mass concentrations used to represent local mass variations when inverting the GRACE derived220
gravity field [e.g. 57, 58], and widely used for estimating glacier mass changes from GRACE221
data [e.g. 59, 60, 61, 16, 62]. The interest of the mascon approach relies on the opportunity to222
incorporate a-priori information on the spatial distribution of mass variability from independent223
sources [e.g. 63]. This has been shown to yield more realistic mass changes [e.g. 19, 64, 65] than224
simple rescaling approaches since the sensitivity kernel of GRACE is not homogeneous over a225
given region but decreases from the centre of the region towards its margins [38]. Using the226
mascon approach for resolving isolated masses such as glaciers or small glacier regions, however,227
poses a tradeoff problem concerning the mascon size: A large number of small mascons with228
uniform mass distribution (the most common assumption for individual mascons) would better229
describe the high spatial variability, but increase the ill-posedness of the inversion problem since230
different distributions of the mass changes can result in the same gravity changes as seen by231
GRACE. The appropriate mascon definition is therefore critical in order to quantify the mass232
changes accurately [e.g. 63].233
In order to determine the optimal mascon repartition, the region of interest was first discretized234
in 1, 3, 7 and 16 different mascons with an equivalent size of about 6.0◦, 3.4◦, 1.8◦, and 1.2◦,235
respectively (not shown). Discretization was performed by discerning individual regions of glacio-236
logical interest. The signal of a known, synthetic mass distribution of 1 Gt in total (Fig. A.8a)237
was then filtered in order to mimic the GRACE resolution and the so obtained synthetic GRACE238
signal re-focussed on the mascons through a Bayesian inversion approach [e.g. 66, 64] in order239
to assess the total uncertainty introduced during the inversion.240
The Bayesian approach consists of propagating the knowledge provided by the GRACE signal241
through the known spatial filtering applied to that signal, and to combine it with the a-priori242
knowledge of the sub-mascon distribution of the actual mass changes that can be gained from243
the known glacier position (see below). Let mobs and mact be the vectors of observed and244
actual mass changes respectively, and f the forward model through which mact is observed (e.g.245
the effect of a spectral truncation at a specific degree and the application of a given Gaussian,246
DDK, or decorrelation filter to the GRACE data; see Section A.2), i.e. mobs = f(mact). The247
a-posteriori probability density function of mact, i.e. p(mact), can be calculated as,248
p(mact) = π(mact) exp
(−1
2rTC−1
MMr
), (1)
where π(mact) (taken as locally uniform) is the a-priori probability density function of mact,249
r = mact − f(mobs) is a vector of residuals, rT is the transpose of r, CMM is the covariance250
matrix of the observations, and C−1MM its inverse. As p(mact) is a probability function, it is251
constrained by∫p(mobs)dm = 1.252
The uncertainty introduced by the glacier representation during the inversion is quantified by253
exploring three different options: In option 1, the spatial distribution of the mass change is254
assumed to be spatially uniform within every individual mascon (Fig. A.8b). In option 2, the255
6
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 7: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/7.jpg)
mass change is assumed to be uniform as well, but is localized at positions where glacier are256
known to occur (Fig. A.8c). To this end, glaciers are represented on a 1/8th degree grid, and257
every grid cell that contains a glacier according to the Randolph Glacier Inventory (RGI [67];258
see also Section C.1) is assumed to be completely glacierized. In option 3, glaciers are again259
represented on a 1/8th degree grid, but the mass change is assumed to be proportional to the260
degree of glacierization (inferred from the RGI) of that grid cell (Fig. A.8d). The resolution of261
1/8th degrees for glacier representation was chosen since tests indicated this grid size yielding262
the best compromise between numerical stability and actual description of the spatial patterns263
at GRACE resolution (not shown). Further a-priori information about glacier mass changes that264
could potentially be gained from ICESat data (see Section B) or glaciological measurements and265
modelling (see Section C) were not considered in order to ensure independence between the three266
approaches. The total uncertainty assessed for the so performed inversion approach is shown in267
Figure A.7.268
The results indicate that in general, a set of three mascons located over the central, eastern and269
northern part of the Tien Shan respectively (Fig. A.8c), is sufficient for an adequate recovery270
of the actual mass change signal: The uncertainty linked to the unknown distribution of glacier271
mass changes is up to 16 % for a truncation of the GRACE signal at degree 120 (leftmost bars in272
Fig. A.7; the filtering corresponds to the one applied for regularized GRACE solutions, and the273
relative high uncertainty reflects the need of correctly quantifying the sub-mascon mass change274
distribution), but below 4 % in all other cases. Thus, for the inversion of the actual GRACE275
signal, three mascons are used over the Tien Shan. Additional mascons are placed over the276
nearby glacierized regions of the Alay Range, the Pamirs, and the Bogda Shan (Fig. A.8c) in277
order to account for leakage from these regions.278
The above described Bayesian approach is computationally expensive when inverting whole279
time series. The approach based on iterative least-square (LS) adjustment proposed by [60]280
was therefore tested as an alternative. In the above test cases, the simple LS inversion scheme281
recovers the synthetic mass distribution within 1% of what recovered by the Bayesian inversion.282
The computationally cheaper LS approach is therefore used for all analyses.283
The quality of the mascon fit is evaluated with the root mean square error of the residuals of the284
LS adjustment (post-fit residuals) and the portion of explained variance. Depending on noise285
structure and processing method, glacier mass changes modelled by mascons explain between286
60 and 95% of the variance in the residual GRACE signal (Fig. A.10). In general, DDK-287
filtered options (Fig. A.10c,f,i) have lowest residuals, followed by stabilized GRGS solutions288
(Fig. A.10j,k). As expected from Figure A.7, in contrast, the higher spatial information of289
the regularized solutions (Fig. A.10l) also translates to higher post-fit residuals. The spatial290
distribution of these residuals, moreover, suggests that glacier mass is indeed changing within291
each of the three sub-regions defined by the mascons, and that glacier mass changes are relatively292
low for regions R3a and R6 (for region definition, see Fig. 5e in the main article). It is additionally293
to note that fitting a limited number of mascons acts as a supplementary filter in the estimation294
process. In fact, any noise which does not have the same spatial structure as imposed by the295
mascon definition (Fig. 2e in the main article), is rejected (Figure A.10). The inclusion of sub-296
mascon mass distributions adds further spatial constrains. Generally speaking, homogeneous297
mascons require larger mass changes than concentrated mass distributions for an equivalent298
quality of the fit.299
The ensemble of time series for the recovered glacier mass anomaly resulting from the combi-300
nation of various GRACE solutions, substitutions of low degree terms, LSM and GIA models301
output (396 options in total) are shown in Figure A.2d. The resulting glacier mass-change rates302
are displayed in Figures 3 and 6 of the main article. The spatial distribution of the GRACE303
residual signal prior and after mascon adjustment is shown in Figures A.9 and A.10, respectively.304
7
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 8: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/8.jpg)
A.6 Uncertainty estimates305
The recovered glacier mass-change rates are affected by a series of uncertainty sources inherited306
from both processing scheme and propagation of model errors. For computing the total error307
budget, all sources are treated as independent.308
Uncertainties for individual GRACE solutions originate mainly from errors in the measurements309
of the satellite range-rate, and errors from imperfectly reduced atmospheric and oceanic mass310
contributions. Root mean square errors (RMSE) for a given month are typically in the range of311
5 mm (10 to 50 mm, depending on latitude; about 15 mm over the Tien Shan) equivalent water312
thickness for the range-rate measurement (the atmospheric and oceanic corrections), and can be313
estimated empirically from residual variations over ocean surfaces [e.g. 68, 69]. Here, GRACE314
TMC uncertainties are estimated as the RMSE over oceans in the same latitude-band as the Tien315
Shan, but excluding ocean regions within 1000 km of continental coast lines to avoid leakage from316
changes in terrestrial water storage [14]. Averaged over the various solutions, GRACE-observed317
TMC for the Tien Shan and the period 2003-2009 are estimated to be −13.3± 3.8 Gt a−1. The318
uncertainty introduced by the GRACE observation (i.e. ±3.8 Gt a−1), thus, induces roughly319
1/2 of the total uncertainty when recovering glacier mass changes (cf. Tab. A.4). The spread320
between individual solutions is slightly larger (±4.6 Gt a−1). This might be linked to the impact321
of the different filtering options, and the destriping filter in particular, which largely affects the322
GRACE signal (Figures A.3). The difference between destriped CSR and GFZ solutions (solu-323
tions G1.p1 and G2.p1, respectively) for example, is 1.3 Gt a−1, while unconstrained GRACE324
solutions filtered with DDK4 yield TMCs that are consistent within 0.5 Gt a−1.325
The variability between individual LSM outputs approximately accounts for the other half of326
the total uncertainty budget (Tab. A.4). The uncertainty introduced in the recovered mass327
change rate, calculated as the mean standard deviation for monthly values of individual model328
outputs, is in the order of 5 Gt a−1. Dispersion among models is clearly scale dependent (cf.329
Fig. A.6 and Section A.3), which hampers any interpretation of trends at the sub-regional scale.330
Moreover, some LSMs clearly depart form the ensemble mean: CLM v2.0 (m01) for example,331
yields negligible hydrological contribution to TMC, whereas the W3RA model (m10), explains332
the totality of the observed TMC trends over the region of interest, thus yielding very small333
glacier change contributions. This clearly highlights the importance of considering an ensemble334
of LSMs driven by different forcing datasets rather than an individual LSM.335
Estimating the uncertainty from estimated SWS changes is difficult. The main uncertainties336
during the period 2003-2009 (period with direct SWS observations; cf. Section A.3) arise from337
(a) the constant-area hypothesis when converting lake level changes to volume changes, and (b)338
the extrapolation of lake levels from nearby lakes. Here, we consider a conservative estimate of339
20%, i.e. an uncertainty of about 0.4 Gt a−1. Note that this estimate would need adjustment340
for the period after 2010.341
The uncertainty in GIA contributions is taken from [15], and estimated to be about 1 Gt a−1.342
According to the results of the GIA models used in the present study (cf. Section A.4) this is343
a very conservative estimate. As described in Section A.4, and in line with [15], uncertainties344
deriving from tectonic and erosion processes are assumed to be negligible.345
All uncertainty sources and their contributions to the estimated glacier mass changes are sum-346
marized in Table A.4. Following the rules of Gaussian error propagation for independent errors347
(cf. Section C.6), the total uncertainty for the glacier mass change rate recovered for the period348
2003-2009 is estimated to be ±6.4 Gt a−1 (95% confidence level).349
A.7 Comparison with previous studies350
So far, only four studies [15, 16, 17, 62] have computed glacier mass change rates for the entire351
Tien Shan based on GRACE data.352
8
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 9: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/9.jpg)
[15] used CSR RL04 data truncated at degree 60 and filtered with a Gaussian smoother of353
150 km half-width (T60 G150), addressed the period 2003-2010, and used LSM outputs from354
NOAH v2.7.1 (our m05) and CLM v4.0 (m02). Using the same GIA correction as these authors355
and the same methodological settings, our estimated average glacier mass change over the stated356
period yields −4.9 ± 3.0 Gt a−1. This is in good agreement with the −5 ± 6 Gt a−1 estimate in357
the original study, and increases the confidence in the developed methodology. The stated358
confidence intervals, however, differ largely. This seems mainly due to the fact that [15] used359
LSMs outputs at full resolution, which, as shown in Figure A.6, is prone to cause a larger spread360
in the estimated trends.361
Focusing on the 2003-2009 time period and using the same methods as above, [16] found glacier362
mass change rates of −7±4 Gt a−1 (pers. comm. A. Gardner, October 2013; results for the Tien363
Shan are not shown explicitly in the original publication but included in the estimates for High364
Mountain Asia as a whole). Also in this case, the results can be reproduced: Our estimate for365
the particular setting yields −7.0± 3.1 Gt a−1, with again, a slightly smaller confidence interval366
probably arising from LSM filtering.367
In a recent study using CSR RL05 solutions, [17] calculated glacier mass change rates for the368
period 2003-2012 by using truncation at degree 60, and a Gaussian smoother with 300 km half-369
width (T60 G300). After the removal of a 5-year cycle that the authors suggested to be linked370
to the Arctic Oscillation and the El Nino-Southern Oscillation, the glacier mass change rate for371
the Tien Shan was estimated to be −8.41 ± 4.82 Gt a−1. Following the same methodology, our372
estimate yields a significantly less negative trend, i.e. −1.4± 0.6 Gt a−1. We suspect this large373
difference to be associated to a combination of three reasons: (i) the use of a destriping filter,374
whose data-adaptive nature makes the filter output hardly predictable, (ii) the low effective375
resolution of the GRACE solutions used by [17] when compared to their mascon size, and (iii)376
the difficulty in exactly reproducing the removal of the mentioned cycle. Concerning point (i)377
Figure A.3a and A.3d show for example, how the spatial distribution of two solutions having378
very similar mean temporal evolution (c.f. solutions G1.p1 and G2.p1 in Fig. A.2) can differ379
substantially when processed with a destriping filter (note for example the movement of about 3◦380
longitude in the centre of the positive trend in the north west corner, or the concentration towards381
the west for the negative trend in the Chinese Tien Shan). The application of more predictable382
decorrelation filters such as DDK (c.f. Figures A.3c, A.3f, and A.3i for example) is therefore383
encouraged. Glacier mass-changes derived from DDK filter are not only more consistent among384
different unconstrained GRACE solutions, but also closer to the estimates by using regularized385
and stabilized GRACE solutions (Fig. A.9).386
[62] is currently the most recent study including our region of interest. In it, a global-scale387
mascon strategy was proposed to infer global ice cap and glacier mass changes for the period388
2003-2013. The glacier mass-change rate for the Tien Shan and the period 2003-2009 was389
estimated to be −10 ± 1.7 Gt a−1 (pers. comm. E.J.O. Schrama, February 2015; similarly as390
in [16], explicit results included in the original publication refer to High Mountain Asia as a391
whole). This is more negative than our estimates (−6.6± 4.0 Gt a−1 when including all options,392
−7.5 ± 4.3 Gt a−1 when excluding Gx.p2-options), but within the stated confidence intervals.393
The discrepancy is explained to a minor extend by the different mascon adjustment approach,394
and mainly by the use of a single LSM (GLDAS NOAH, our model m05a in Tab. A.2) in their395
study. Indeed, for the region of interest, the particular model predicts a less negative trend396
in SMS and SPP as compared to alternate models (Fig. 2c of the main article). When using397
LSM-options m05a only, our estimates reach −7.9±2.3 Gt a−1 (including all remaining options)398
and −8.5 ± 1.9 Gt a−1 (when excluding remaining Gx.p2-options), respectively, thus being in399
satisfactory agreement.400
9
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 10: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/10.jpg)
FReal world: TMC = SWS + SMS + GWS + SSP + GIA + ICE
GRACE resolution: TMCF = SWSF + SMSF + GWSF + SSPF + GIAF + ICEF
Spatial �ltering (straightforward mathematical process)
Mascon approach (includes a -priori information)
Green = Constrained from measurementsBlue = Constrained from model results
Black = Non-observed quantitiesRed = Target quantity
②
③①
Processing �ow
Figure A.1: Schematic representation of the approach used for the GRACE-derived estimates. Observedmass change contribution from surface water storage (SWS), and modelled contributions from soil mois-ture storage (SMS), groundwater storage (GWS), and the seasonal snow pack (SSP) are filtered in orderto achieve the same spatial resolution as the total mass change (TMC) observed by GRACE (¬). Allfiltered mass contributions (subscript F ), including the effect of glacial isostatic adjustment (GIA), areremoved from the observed TMC (). The obtained residual is interpreted as contribution from glaciers(ICE) and back-transformed to real-world resolution by using a mascon approach (®).
10
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 11: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/11.jpg)
-60
-40
-20
0
20
40
60O
bser
ved
TM
C a
nom
aly
( Gt )
G1.p1G2.p1
G3.p1G4.p1
G5.p0G6.p0
G7.p0GRACE solutions
a
-60
-40
-20
0
20
40
60
Mod
elle
d T
WS
ano
mal
y ( G
t )
m01m02
m03m04
m05m05a
m06m07
m08m09
m10LSMs
c
-8
-6
-4
-2
0
2
4
6
Filt
ered
SW
S a
nom
aly
( Gt )
noneT120
T60T50
T60 G150T60 G300
T60 DDK4Applied filters
b
2004 2006 2008 2010 2012
-60
-40
-20
0
20
40
60
Gla
cier
mas
s an
omal
y ( G
t ) d
Figure A.2: Anomalies in (a) total mass storage (TMC) as observed by GRACE, (b) surface waterstorage (SWS) contributions to TMC, (c) total water storage (TWS) as modelled with the suite of landsurface models (LSMs) given in Table A.2, and (d) recovered total glacier mass over the Tien Shanfor all used processing options. Anomalies are computed with respect to the average over the period2003-2009 and refer to the region defined in Figure A.8 (437 000 km2). For the nomenclature of thedifferent GRACE solutions refer to Table A.1. SWS data are available for the period 2003-2010 only, anda constant anomaly is assumed outside that period. For the filtering methods, Tn denotes the truncationat degree and order n, Gw a Gaussian smoother with half-width w (km), and DDK4 a decorrelation filterafter [29]. The red line in panel (d) is the 2003-2009 mass-change trend. Note that in (b) the scale differsfrom the other panels.
11
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 12: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/12.jpg)
ihg
fed
cba
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
Mea
n TM
C tr
end
(Gt a
-1)
10-2
020
-10
0
G 4 . p 0 G 5 . p 0 G 6 . p 0
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
G1
.pX
G2
.pX
G3
.pX
G x . p 1 G x . p 2 G x . p 3
lkj
Figure A.3: Mean trend of total mass change (TMC) for (a-i) unconstrained, (j-k) stabilized, and (l)regularized GRACE solutions. The nomenclature on the right-hand-side and figure top is according toTable A.1. X, and x stand for an arbitrary processing method or centre, respectively. Trends refer tothe period 2003-2009.
12
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 13: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/13.jpg)
fT60 G300
dT60 DDK4
bT60 G150
eT50
cT60
aT120
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
72°E 76°E 80°E 84°E 88°E72°E 76°E 80°E 84°E 88°E
Mean SWS trend (Gt a-1)
10-20 20-10 0
Figure A.4: Spatial distribution of the surface water storage (SWS) contribution to the total mass-changetrend during the period 2003-2009 as seen at GRACE resolution. Individual panels refer to individualfiltering options. The filter notation (upper-left corner) is according to Fig. A.2b. Outlines of lakes (givenin red) that are (not) accounted for are (not) filled.
13
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 14: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/14.jpg)
ic46°N
44°N
40°N
38°N
42°N
h
g
b46°N
44°N
40°N
38°N
42°N
a46°N
44°N
40°N
38°N
42°N
Mean TWS trend (Gt a-1)
10-20 20-10 0
Standard deviation (Gt a-1)
0 105
T120T60
T50
l
72°E 76°E 80°E 84°E 88°E
f46°N
44°N
40°N
38°N
42°N
72°E 76°E 80°E 84°E 88°E
k
j
e46°N
44°N
40°N
38°N
42°N
d46°N
44°N
40°N
38°N
42°N
T60 G150
T60 DD
K4
T60 G300
Figure A.5: Spatial distribution of the (a-f) mean trend and (g-l) according standard deviation for thetotal water storage (TWS) changes predicted by the ensemble of land surface models given in Table A.2.Individual panels refer to individual filtering options (filter notation on the right-hand side according toFig. A.2b). Values refer to the period 2003-2009.
14
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 15: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/15.jpg)
-8
-6
-4
-2
0
TWS
tren
d (G
t a-1)
Full resolution T120 T60 T50 T60G150 T60G300DDK4T60
Figure A.6: Effect of various filtering options (notation according to Fig. A.2c) applied to the ensembleof land surface models (LSMs) given in Table A.2. The boxplots show the spread between minimum andmaximum values (whiskers), the interquartile range (box), and the average value (horizontal line insidethe box) for the 2003-2009 trend in total water storage (TWS) as calculated by the various LSMs. Thedegree of filtering increases from left to right. Note the reduction in spread for increasing filtering degree.
0
5
10
15
20
Tota
l unc
erta
inty
(%
)
T120 T60 T50 T60G150 DDK4 T60G300
1 mascon, 6.0°3 mascons, 3.4°7 mascons, 1.8°
16 mascons, 1.2°
T60
Figure A.7: Total uncertainty introduced during the mascon inversion (step ® in Figure A.1) as a func-tion of mascon size and filtering option. The notation for the filtering options is according to Fig. A.2b.
15
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 16: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/16.jpg)
ba
0 0.40.2 0.6 0.8
Synthetic mass (Gt)
46°N
44°N
40°N
38°N
42°N
d
72°E 76°E 80°E 84°E 88°E
0 5 10 15%
Glacierization
c
Alay Range and Pamirs
central part
northern part
western partBogda Shan
72°E 76°E 80°E 84°E 88°E
46°N
44°N
40°N
38°N
42°N
Figure A.8: (a) Synthetic mass distribution used for the assessment described in Section A.5, and (b-d)various options assumed for sub-mascon mass change distribution: (b) uniform mass change over theindividual mascons, (c) uniform mass change concentrated in the highlighted pixels, (d) mass changedistribution proportional to the glacierized area within individual pixels. In all panels, grey outlinesshow glacierized areas according to the RGI v3.2. Panel (c) gives the names of the individual mascons asused in Section A.5.
16
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 17: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/17.jpg)
ihg
fed
cba
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
Mea
n tre
nd (G
t a-1)
10-2
020
-10
0
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
G1
.pX
G2
.pX
G3
.pX
G x . p 1 G x . p 2 G x . p 3
lkj
G 4 . p 0 G 5 . p 0 G 6 . p 0
Figure A.9: Mean trend of recovered glacier mass-change rates prior to mascon adjustment for (a-i)unconstrained, (j-k) stabilized, and (l) regularized GRACE solutions. The nomenclature on the right-hand-side and figure top is according to Table A.1. X, and x stand for an arbitrary processing methodor centre, respectively. The mean refers to the ensemble of different land surface models (cf. Table A.2)and the period 2003-2009.
17
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 18: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/18.jpg)
ihg
fed
cba
lkj
Mea
n tre
nd (G
t a-1)
10-2
020
-10
0
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
46°N
44°N
40°N
38°N
42°N
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
72°E
76°E
80°E
84°E
88°E
G1
.pX
G2
.pX
G3
.pX
G x . p 1 G x . p 2 G x . p 3
G 4 . p 0 G 5 . p 0 G 6 . p 0
RM
SE
9.3
mm
E.V
ar.
72
%
RM
SE
3.9
mm
E.V
ar.
98
%
RM
SE
4.3
mm
E.V
ar.
87
%
RM
SE
9.5
mm
E.V
ar.
61
%
RM
SE
3.5
mm
E.V
ar.
98
%
RM
SE
4.5
mm
E.V
ar.
86
%
RM
SE
8.6
mm
E.V
ar.
76
%
RM
SE
3.9
mm
E.V
ar.
95
%
RM
SE
4.1
mm
E.V
ar.
87
%
RM
SE
4.5
mm
E.V
ar.
95
%
RM
SE
7.6
mm
E.V
ar.
67
%
RM
SE
6.4
mm
E.V
ar.
88
%
Figure A.10: Residual signal after mascon adjustment (post-fit residuals) for (a-i) unconstrained, (j-k)stabilized, and (l) regularized GRACE solutions. The root mean square error (RMSE) of the fit and theportion of explained variance (E.Var) are provided. The nomenclature on the right-hand-side and figuretop is according to Table A.1. X, and x stand for an arbitrary processing method or centre, respectively.The mean refers to the ensemble of different land surface models (cf. Table A.2) and the period 2003-2009.
18
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 19: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/19.jpg)
Table A.1: Overview of the different GRACE solutions and processing strategies used in this study.Option is a code which, combined with the codes in Table A.2, is used for distinguishing the variousestimates. The first (second) part of the code defines the GRACE solution (processing method), whereX is either 1, 2, or 3. RL is the product release, dmax is the maximal degree and order provided inthe Level-2 data, TI is the time integration of the solution, and Ref. a reference for the data source.In the column Processing, T dn indicates truncation at a particular degree n, and D, Gw, and DDK4indicate the application of a decorrelation filter after [27], a Gaussian filter with half-width w (km), or adecorrelation filter after [29], respectively.
Option Solution RL Type dmax TI Processing Ref.
G1.pX CSR 5 Unconstrained 60 30 d either X=1: T dmax + D + G150 [70]G2.pX GFZ 5a Unconstrained 90 30 d or X=2: T dmax + D + G300 [71]G3.pX JPL 5 Unconstrained 90 30 d or X=3: T dmax + DDK4 [72]
G4.p0 GRGS 2 Stabilized 50 10 d T dmax [32]G5.p0 GRGS 3 Stabilized 80 30 d T d60 [33]
G6.p0 CSR reg. 3 Regularized 120 30 d T dmax [25]
Table A.2: Overview of the different land surface models used in this study. Option is the code usedin combination with the codes in Table A.1 for distinguishing various estimates. Ver. indicates themodel version, Res. is the spatial resolution of the model output (irr. stands for irregular), and Ref. isa reference for the given model. The column Represented compartments indicates which water storagecompartments are represented in the models; SWS, SMS, GWS, and SSP are the storages in surface water,soil moisture, groundwater, and the seasonal snow pack, respectively; ant. indicates that anthropogenicwater use is represented as well. Forcing indicates which climatic forcing is used as model driver.
Option Model Ver. Res. Represented compartments Forcing Ref.
m01 CLM (GLDAS) 2 1◦ SMS, SSP CMAP + NOAA/GDAS [73]m02 CLM 4 irr. SWS, SMS, GWS, SSP GPCP + CRUNCEP [74]m03 CLM 4.5 irr. SWS, SMS, GWS, SSP, ant. GPCP + CRUNCEP [75]m04 MOSAIC (GLDAS) - 1◦ SMS, SSP CMAP + NOAA/GDAS [73]m05 NOAH (GLDAS) 2.7.1 0.25◦ SMS, SSP CMAP + NOAA/GDAS [73]m05a NOAH (GLDAS) 2.7.1 1◦ SMS, SSP CMAP + NOAA/GDAS [73]m06 NOAH (GLDAS-2) 3.3 1◦ SMS, SSP Princeton University [76] [73]m07 VIC (GLDAS) - 1◦ SMS, SSP CMAP + NOAA/GDAS [73]m08 WGHM 2.1 0.5◦ SWS, SMS, GWS, SSP, ant. GPCC + ECMWF [77]m09 WGHM 2.2 0.5◦ SWS, SMS, GWS, SSP, ant. GPCC + ECMWF [78]m10 W3RA - 1◦ SMS, GWS, SSP Princeton University [76] [79]
19
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 20: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/20.jpg)
Table A.3: Mass changes for major lakes and reservoirs close to the study area. The stated trendsand average seasonal variations (SV) refer to the period 2003-2009, and are given in Gt a−1 and Gt,respectively. Meth. indicates the method used for estimating the change (V = from reported lake volume;LL = from reported lake level; E = estimated from nearby records; see Section A.3 for details), and Ref.is a reference for the data source. Given coordinates are approximate. Lakes and reservoirs are listedaccording to the trend magnitude.
Name Type Coordinates Obs. SV Trend Ref.
Zaysan Reservoir 48.00 N 84.00 E LL 2.123 -1.415 Hydroweb1
Toktogul Reservoir 41.78 N 72.83 E V 1.317 -1.105 HydrowebBalkhash Lake 46.10 N 74.10 E LL 3.583 -0.825 HydrowebBosten Lake 41.98 N 87.00 E V 0.259 -0.454 HydrowebKayrakkum Reservoir 40.30 N 70.10 E E 1.164 -0.109 -Tchardarin Reservoir 41.13 N 68.13 E LL 2.243 -0.094 HydrowebKapchagay Reservoir 43.80 N 77.50 E LL 0.234 -0.072 HydrowebUlungur Lake 47.25 N 87.20 E LL 0.097 -0.057 HydrowebAlakol Lake 46.15 N 81.65 E E 0.530 -0.041 -Jili Lake 46.91 N 87.45 E E 0.011 -0.012 -Sasykkol Lake 46.55 N 80.95 E LL 0.139 -0.011 HydrowebEbinur Lake 44.88 N 82.92 E E 0.134 -0.010 -Sayram Lake 44.60 N 81.20 E E 0.086 -0.007 -Uyaly Lake 46.44 N 81.28 E E 0.021 -0.002 -Sarez Lake 38.20 N 72.78 E E 0.049 0.002 -Karakul Lake 39.03 N 73.40 E E 0.187 0.006 -Charvak Reservoir 41.65 N 70.03 E V 0.295 0.014 SIC-ICWC2
Issyk Kul Lake 42.40 N 77.30 E LL 0.622 0.050 HydrowebAydarkul Reservoir 40.95 N 66.50 E LL 1.009 0.102 Hydroweb
1 LEGOS Hydroweb (www.legos.obs-mip.fr/soa/hydrologie/hydroweb/) [40]2 Scientific-Information Center of the Interstate Coordination Water Commission
of the Central Asia (www.sic.icwc-aral.uz)
Table A.4: Summary of the contributions to total mass changes (TMC) observed by GRACE andcorresponding uncertainty sources. Values (Gt a−1) and uncertainties (Uncert.; 95% confidence level) areaverages and refer to the period 2003-2009. TWS = total water storage; SWS = surface water storage;neglig. = negligible.
Source Value Uncert. Notes
GRACE TMC rate −13.3 3.8Substitution of degree 1 and 2 terms < 0.1 0.1
Unknown actual spatial distribution of TMC 0 ∼ 0.7 (1)(2)
TWS correction (excluding SWS) +3.6 5.0 (1)
SWS correction +2.1 0.4 (1)(3)
Correction for tectonic uplift and erosion neglig. neglig.
GIA correction +1 1 (4)
LIA correction 0 0 (4)
Recovered glacier mass change -6.6 4.0 (5)(6)
(1) Depending on filtering option(2) 4-16% of the recovered glacier mass change (see Figure A.7)(3) 20% of the correction (4) Based on [15](5) −7.5 ± 4.3 Gt a−1 when excluding Gx.p2 options(6) Confidece interval corresponds to 2 times the standard deviation of all ensemble members
20
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 21: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/21.jpg)
B Estimates based on ICESat data401
As part of NASA’s Earth Observing System of satellites, the Ice, Cloud and Land Elevation402
Satellite (ICESat) mission was operative between February 2003 and October 2009 for a total of403
18 observation campaigns lasting between 12 and 55 days [80]. During the individual campaigns,404
the onboard Geoscience Laser Altimeter System (GLAS) measured time delays between 1064 nm405
laser pulse transmissions and surface echo returns, thus deriving time series of surface elevation406
along its repeat-track orbits [81]. GLAS ground footprints were separated by about 170 m407
along-track, and had a varying elliptical shape with an average diameter of about 70 m [82, 83].408
Although originally designed for measuring elevation changes of the Greenland and Antarctic409
ice sheets, GLAS measurements have successfully been applied for deriving mass changes of410
glaciers outside the polar regions, such as in the Himalayas [84], High Mountain Asia [16], or411
the Tibetan Plateau [85]. Here, GLAS/ICESat L1B Global Elevation Data (GLA06) altimetry412
product release 33 [86] is used. For calculating glacier volume changes, three different approaches413
presented by [16] are reproduced and extended. In particular, the elevation dependency that414
can be detected in the elevation change signal for the glaciers in the Tien Shan (see also the415
supplementary material by [16]) is explicitly accounted for in the spatial extrapolation.416
B.1 Basic pre-processing and auxiliary data417
Because of the relatively large cross-track separation distances of ICESat orbits at lower latitudes418
(distances up to a few kilometres are common), pure repeat-track methods that are commonly419
used in polar regions [e.g 87, 88]) are not suitable for the region of interest. To overcome420
the problem and targeting the Hindu Kush-Karakoram-Himalaya area, [84] proposed to use421
an external DEM to correct for topographic differences between the altimetry measurements,422
and to analyse elevation trends within the so-corrected ICESat data for deriving mass budgets.423
The same method was used for entire High Mountain Asia by [16], who additionally analysed424
elevation differences between nearest-neighbour points in a slightly modified version of the DEM425
projection-method of [89]. The here presented study, relies on and extends the methods described426
in [16].427
All analyses are performed by using orthometric elevations in the Universal Transverse Mer-428
cator (UTM) projection of WGS84. These are obtained by subtracting the geoid of the Earth429
Gravity Model 2008 [90] from the ellipsoidal elevations of the GLA06 product, and converting430
the default TOPEX/Poseidon datum [91] into WGS84. Saturated return waveforms, induced431
by detector overload from return energy of near-specular reflectors [92], are corrected by adding432
the correction product provided jointly to the data set.433
The void-filled version 4 of the Shuttle Radar Topography Mission (SRTM) DEM provided by434
the Consultative Group on International Agricultural Research [93] was used as the reference435
DEM. [16] showed that for retrieving surface elevation changes with the proposed methods, the436
SRTM DEM is better suited than the potentially competing Global DEM obtained from the437
Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER GDEM). Part of438
the reason is that the SRTM DEM refers to a single point in time (February 2000), whereas the439
ASTER GDEM is constructed by stacking and averaging stereoscopic imagery acquired between440
the years 2000 and 2011.441
Following [16], DEM elevation and slope for each ICESat footprint is extracted from the reference442
DEM through bi-linear interpolation of the DEM grid cells. Glacier surfaces are discerned by443
using the glacier outlines provided by the RGI (see Section C.1). ICESat footprints showing444
a difference larger than ±100 m from the reference DEM, as well as footprints over void-filled445
SRTM grid cells are removed from further analysis. This results in the removal of 10% of the446
data, and a remaining sample of about 15,000 ICESat footprints over glaciers (Fig. 4a of the447
main article).448
Histogram analysis reveals that the so-reduced sample is representative for the glacierized sur-449
21
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 22: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/22.jpg)
faces in the region of interest in terms of both spatial distribution and morphological characteris-450
tics (Fig. B.1). A tendency in over-representing flat and steep slopes can be detected (Fig. B.1f).451
Note, however, that in this case local values from the ICESat footprints (grey line) are compared452
to glacier-averaged values (red line), and that assigning the corresponding glacier-averaged value453
to the individual footprints (dashed line) eliminates the difference.454
The temporal consistency of the elevations retrieved by ICESat in the region of interest is455
assessed over non-glacierized terrain. Following [16] and [84], footprints over land areas within456
a 5 km buffer from glaciers are considered, and the data are decimated in order to achieve457
agreement between the slope histograms for glacierized and non-glacierized terrain. Robust458
trend fitting through the so obtained data sample of about 14,400 data points reveals a non-459
significant trend of 0.03± 0.07 m a−1 (dark grey line in Fig. 4b of the main article), which is in460
good agreement with the values found for the entire region of High Mountain Asia [84, 16]. The461
analysis also reveals a systematic elevation difference of about 1.4 m between ICESat and SRTM462
DEM (Fig. 4b of the main article). This has previously been explained with the low-frequency463
biases in the SRTM elevations [94], and the amplitude of the bias is in line with the analyses by464
[84] (see their supplementary Table S2 for example). Note that since all methods described in465
the following consider either elevation differences (method I1) or elevation trends (methods I2466
and I3), the detected bias has no influence on the derived elevation change rates.467
B.2 Method I1 : Elevation difference between nearby footprints468
Method I1 corresponds to method “A” in the analysis of High Mountain Asia by [16]. Basically,469
the method estimates local elevation change rates by dividing the difference in elevation between470
two nearby ICESat footprints with the time span between the two acquisition dates. Since the471
footprints are not at the same location, the ICESat elevations are first corrected for topographic472
differences by subtracting the elevation of the reference DEM at the particular location. Ele-473
vation differences are only computed between footprints stemming from the same season and474
separated by 3 or more integer years (i.e. autumn-to-autumn or winter-to-winter comparisons).475
The comparison between the same season allows to minimize the effects of the temporally vary-476
ing snow cover, whereas a separation by 3 or more years is necessary for achieving a sufficiently477
large signal-to-noise ratio [16]. The obtained elevation change rates are then averaged within478
clusters of 5 km radius in order to reduce the potential bias due to uneven spatial sampling479
and the large spacing between ICESat tracks. Following [16], an iterative 3-standard-deviation480
outlier removal (5% convergence) was applied to the data prior to averaging in order to re-481
duce the sensitivity to gross errors. For the analyses, all autumn and winter ICESat data from482
campaign L2a onwards (autumn 2003 and later) were used, i.e. only data from the calibration483
campaign (L1) and the three summer campaigns (L2c, L3c, L3f) were discarded (Fig. 4a of the484
main article). This resulted in a total of about 180 clusters, each composed of approximately485
80 footprints.486
B.3 Method I2 and I3 : Elevation trends within ICESat data487
Methods I2 and I3 correspond to methods “B” and “C” of [16], respectively, which are in488
turn based on the ideas by [84]. Both methods estimate the elevation change rate within a re-489
gion by computing the temporal trend in the difference between all ICESat footprints available490
for that region and the reference elevation extracted from the reference DEM. The difference491
between method I2 and I3 is given by the considered ICESat campaigns: Method I2 (I3) con-492
siders only data collected during autumn (winter) campaigns (Fig. 4b of the main article). As493
in [16], glacierized surfaces are not subdivided further into debris-free and debris-covered ice,494
since the differences have been shown to be not significant [84, 95]. Trend fitting is performed495
through (a) ordinary least squares fitting, after a 3-standard-deviation edit for outlier removal496
(methods I2.a and I3.a), or (b) robust linear regression [e.g. 96] with iterative weighting based on497
22
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 23: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/23.jpg)
Tukey’s biweight function (e.g. [97]) (methods I2.b and I3.b). In both cases, constant campaign-498
wise weights were additionally assigned in order to account for the unequal number of ICESat499
footprints within individual campaigns. The analyses include ICESat data collected between500
campaigns L2a (autumn 2003) and L2e (winter 2009). Data from the calibration campaign L1501
(winter 2003), the incomplete campaign L2f (autumn 2009), and the three summer campaigns502
(L2c, L3c, L3f) were discarded (Fig. 4a-b of the main article), as discussed by [84].503
B.4 Regional extrapolation and volume to mass conversion504
The sparse ICESat footprint coverage and the relatively high level of noise in the calculated505
elevation change rates hampers a robust spatial interpretation of the results. Compared to506
other regions in High Mountain Asia, a pronounced correlation between elevation change rate and507
altitude can be found for the glacierized surfaces in the Tien Shan (see Supplementary Figure S8508
in [16]). Here, two different options (options e2 and e3) are proposed for utilizing this relation509
in the spatial extrapolation. These options are compared to the “benchmark solution” (option510
e1) which simply applies the average elevation change rate to all glacierized surfaces [84, 16].511
The altitudinal dependence of the elevation change rate is determined as follows (Fig. 4d of the512
main article): Methods I1, I2, and I3 (including the options I2.a, I2.b and I3.a, I3.b) are applied513
individually to subsamples of ICESat footprints selected according to the local surface elevation514
given by the reference DEM. The subsamples are chosen to contain all footprints within ±100 m515
of a given altitude z, and the procedure is repeated by systematically varying z in uniform steps516
of 1 m within the altitude range of the study region. For methods of the group I1 (I2 and I3),517
the altitude-specific elevation change rate and the according standard error is estimated if at518
least 5 ICESat clusters (50 ICESat footprints) are available in the considered elevation band519
(Fig. 4c of the main article). The so obtained function is then either approximated through a520
linear relation obtained by robust regression (option e2) or through a non-parametric relation521
obtained by smoothing the function with a running median of 200 m elevation width (option522
e3). In both cases, the extrapolation outside the domain within which a change rate could523
be estimated is performed by assuming a constant value. As an example, Figure 4d of the524
main article visualizes the various altitude-dependent functions obtained when using methods525
I2.b. The interval of 200 m used for selecting subsamples of ICESat footprints and smoothing526
the altitude-dependent function was determined empirically, and is a trade-off between a high527
resolution of elevation and a sufficiently large sample for meaningful parameter estimation.528
When determining the elevation change rate of a given glacier within the study region, the529
median elevation of the glacier is computed by intersecting the glacier outlines provided by530
the RGI with the reference DEM, and the elevation change rate determined for that altitude531
is assigned. The intersection of the three options (e1, e2, and e3) with the different methods532
described in the previous section (methods I1, I2.a, I2.b, I3.a, and I3.b) give rise to a total of533
15 combinations for the performed extrapolations (see Tab. B.1).534
No reliable information exist about firn compaction rates in the study area. A recent study by535
[98], however, was able to show that “assuming a value of 850±60 kg m−3 to convert volume536
change to mass change is appropriate for a wide range of conditions”. The proposed value is537
adopted. For a more detailed discussion on the topic, refer to [98].538
The mass change rates resulting for the entire study region from the 15 options are shown in539
Figure 3b of the main article and given in the last column of Table B.1.540
B.5 Uncertainty estimates541
Several sources of uncertainty affect the total error budget for the mass change estimates based542
on ICESat data. Following former works [e.g. 99, 84, 16] and the notation introduced by [99],543
the total uncertainty is estimated as544
σMB = (σ2STDE + σ2
BIAS + σ2SPAT + σ2
TEMP + σ2AREA + σ2
DENS)1/2, (2)
23
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 24: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/24.jpg)
where σMB is the standard deviation of the estimated total mass change of the region of interest;545
σSTDE is the standard error of the coefficients of the relations fitted with the various options546
given in Table B.1 and is an expression of (a) random observational errors in the ICESat data547
[84] and (b) the capability of the fitted functions in reproducing the altitudinal dependency548
of the elevation change rates; σBIAS accounts for the uncertainty introduced by not correcting549
for unknown systematic offsets due to inter-campaign biases [100] or crustal uplift [101]; σSPAT550
takes into consideration potential systematic spatial biases that can occur because of uneven551
sampling of different terrain characteristics (e.g. over-representation of flat slopes, Fig. B.1f);552
σTEMP accounts for potential biases related to uneven temporal sampling (e.g. campaigns with553
less data because of particularly cloudy conditions); σAREA is the uncertainty introduced by the554
imprecisely known glacier area; and σDENS is the standard error associated with the density555
assumed for converting volume changes into mass changes. σSTDE is option specific, and has a556
value ranging between 0.06 and 0.19 m a−1 for options I2.b.e3 and I1.a.e1, respectively. σBIAS,557
σSPAT, and σTEMP are all three set to 0.06 m a−1, following [16]. σAREA is assumed to be 20%558
(a more conservative estimate than assumed in [84, 16] for example), and σDENS is computed559
from the standard error of the density given by [98].560
Note that (1) since the distinction between footprints on glacierized and on non-glacierized561
terrain is performed trough the glacier outlines provided by the RGI (which generally refers to562
the pre-ICESat era; see Section C.1), ice thickness change rates at very low elevations (i.e. at563
elevations where glacier retreat is most pronounced) can potentially be underestimated (since564
footprints erroneously classified as “on glacier” are expected to show change rates close to zero),565
and (2) the misclassification described in (1) has a negligible effect on the overall mass balance566
estimate since the regional extrapolation is performed with the same glacier inventory.567
B.6 Comparison with previous studies568
To date, the only study computing glacier elevation changes from ICESat data in the Tien569
Shan is the one by [16]. The methodologically similar work by [84] excluded the mountain570
range and focused on the Hindu Kush-Karakoram-Himalaya region only. The here presented571
options I1.a.e1, I2.b.e1, and I3.b.e1 are identical with the methods “A”, “B”, and “C” of [16],572
respectively, thus allowing to cross validate the implementation of the methods. As expected, the573
results agree well. The deviations in the mean values (−1.3 Gt a−1, −0.4 Gt a−1, and −0.7 Gt a−1574
for methods “A”,”B”,”C”, respectively; negative values indicating that less negative mass change575
rates are calculated in the here presented study) reflect the differences in (a) assumed density576
for the volume to mass change conversion [850 kg m−3 in this study, and 900 kg m−3 in 16], and577
(b) slight differences in the study region (here, the Bodga Shan range, located north east of578
Urumqui Glacier No.1, was not included) and glacier inventory (RGI v3.2, vs RGI v2.0), which579
jointly result in a difference in glacierized area of about 800 km2 [13,700 km2 in this study, and580
14,500 km2 in the analyses by 16].581
24
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 25: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/25.jpg)
Table B.1: Overview of the 15 different model options for estimates based on ICESat data. The accordingmass change rate dM/dt for the entire study area is given in Gt a−1. Confidence intervals refer to the 95%level. For the average, the stated confidence interval corresponds to two times the standard deviationof all ensemble members. Abbreviations: Elev. = Elevation; diff. btw. = difference between; LS = LeastSquares; regr. = regression.
Code Method Fitting Elev. dependence Period dM/dt(Gt a−1)
I1.a.e1 None 2003-2009 −8.97 ± 3.13I1.a.e2 Elev. diff. btw. nearby footprints Linear 2003-2009 −6.33 ± 2.65I1.a.e3 Non-parametric 2003-2009 −5.88 ± 1.95
I2.a.e1 None 2003-2008 −4.60 ± 2.29I2.a.e2 Ordinary LS Linear 2003-2008 −3.97 ± 1.92I2.a.e3 Non-parametric 2003-2008 −3.93 ± 1.77
Elev. trends within autumn dataI2.b.e1 None 2003-2008 −4.54 ± 2.30I2.b.e2 Robust regr. Linear 2003-2008 −3.89 ± 1.88I2.b.e3 Non-parametric 2003-2008 −3.85 ± 1.70
I3.a.e1 None 2003-2009 −6.59 ± 2.28I3.a.e2 Ordinary LS Linear 2003-2009 −4.49 ± 1.83I3.a.e3 Non-parametric 2003-2009 −5.73 ± 1.83
Elev. trends within winter dataI3.b.e1 None 2003-2009 −6.84 ± 2.28I3.b.e2 Robust regr. Linear 2003-2009 −4.87 ± 1.94I3.b.e3 Non-parametric 2003-2009 −5.82 ± 1.98
AVERAGE 2003-2009 −5.35 ± 2.88
25
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 26: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/26.jpg)
75 80 850
1
2
3
4
5 longitude adegrees east
40 41 42 43 44 45
0
1
2
3
4
5latitude bdegrees north
-2 -1 0 1 20
1
2
3
4
5 log(area) clog(km2)
1.5 2.0 2.5 3.0 3.5
0
1
2
3
4
5log(elev. range) dlog(m)
3000 4000 5000 60000
1
2
3
4
5 mean elev. em a.s.l.
0.8 1.0 1.2 1.4 1.6
0
1
2
3
4
5log(mean slope) flog(degrees)
100 200 3000
1
2
3
4
5 mean aspect fdegrees from north
-2 -1 0 1 2
0
1
2
3
4
5log(pot. radiation) hlog(W m-2)
RGI outlines
Monitored glaciers
ICESat data
Figure B.1: Smoothed histograms of various morphological characteristics for the glacierized areas inthe study region. Characteristics for all glacierized surfaces included in the RGI (red) and for consideredICESat footprints (grey) are shown. Black dots represent the seven glaciers for which mass balance timeseries are available (see Tab. C.1). Units of the abscissa are given below the panel description (upper leftcorner), whilst the ordinate is histogram density given in % (the area under each curve sums up to 100%).Apart from panels (c) and (d), values for glaciers are averaged over the according glacier area, whilstvalues for ICESat footprints are averaged over the footprint extent. In (c) and (d) the value referring tothe entire glacier is assigned to each ICESat footprint individually. In panel (f) the dashed histogram isobtained by assigning the mean glacier slope to the according ICESat footprint. Note that the variablesin panels (c), (d), (f), and (h) are logarithmically transformed. Potential solar radiation (h) is computedas described in Section C.3.
26
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 27: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/27.jpg)
C Estimates based on glaciological modelling582
Similarly as for the estimates based on GRACE and ICESat data, the estimates based on583
glaciological modelling explore a range of different possibilities. This range comprehends a total584
of 30 model options, and can mainly be subdivided into two groups: The first group (M1) relies585
on actual field measurements only, whilst the second (M2) makes additional use of mass balance586
modelling. The two groups are described separately in the next sections.587
C.1 Available glaciological measurements and glacier outlines588
For Central Asia, in-situ mass balance observations (i.e. observations derived from repeated589
stake readings) with a length of ≥ 5 years are available for seven glaciers (Table C.1, Fig. 1 of590
the main article). Out of these series, only two (the series for Tuyuksu Glacier and Urumqi591
Glacier No.1) are still maintained at present. All time series comprehend measurements of both,592
annual mass balance and accumulation. Geodetic glacier ice volume changes are available for593
three selected regions (see Fig. 1 of the main article) and different time periods (Table C.1).594
Sparse additional information on glacier mass balance exists for 13 other glaciers. These time595
series, however, cover either a very short period (< 4 years), or stem from very small glaciers596
(area <1.5 km2) that are adjacent to Tuyuksu Glacier, and are thus not further considered.597
For characterizing the glaciers in the study region individually, the glacier outlines provided by598
the Randolph Glacier Inventory (RGI) version 3.2 [67, 102] are used. In the region of interest,599
large parts of the RGI are taken from the database of the Global Land Ice Measurements from600
Space (GLIMS) initiative [103]. For the Chinese part of the region, RGI data were obtained601
from the first Chinese glacier inventory [104] representing the 1970s and 1980s. For most of the602
Kyrgyz and Kazakh part of the Tien Shan, RGI outlines were mapped semi-automatically from603
ASTER and Landsat TM/ETM+ scenes referring to the period 1999-2003 [105, 106, 107, 108].604
Minor missing areas in western Kyrgyzstan were filled by the outlines compiled by [109] from the605
Digital Chart of the World [110], or the World Glacier Inventory [111]. The uncertainty deriving606
from the imprecise temporal attribution and the uneven quality of the inventory is included into607
the final uncertainty estimate (section C.6). The outlines for Abramov and Shumskiy Glacier608
(Fig. 1 of the main article), known to be inaccurate in RGI v3.2, are replaced with outlines609
digitized manually from 2007 satellite imagery. Basic characteristics of the glacierized surfaces610
(e.g. distributions of size, elevation, aspect, slope, etc.) are shown in Figures B.1 and C.2.611
A number of studies have addressed glacier area changes in the region of interest. These studies612
are summarized in Table C.2 and the information is used for prescribing glacier area changes613
within all of the considered model options (section C.5).614
C.2 M1 : Estimates based on measurements only615
Estimates of the group M1 are exclusively based on in-situ measurements. The general idea is616
to first generate a continuous time series covering the period 1961-2012 for all seven glaciers617
with observations ≥ 5 years, and then to extrapolate those measurements in space. Four dif-618
ferent methods are explored. The first two methods (M1.a and M1.b) are taken from [112];619
the remaining two methods (M1.c, M1.d) follow the same principle but additionally use the620
information that is available from the two glaciers monitored continuously. All estimates refer621
to annual balances. Accumulation measurements are not considered.622
Let bi(t) be the annual balance of glacier i for year t, and let ty (tn) denote years in which the623
mass balance was (not) measured. The different options can be described as follows:624
M1.a: Constant rate625
For each glacier with an incomplete time series, the annual mass balance in years without626
measurements is set to the average balance of the period in which measurements were available627
27
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 28: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/28.jpg)
for this glacier [112]:628
bi(t) = bi(ty), (3)
where t ∈ tn and X is the arithmetic average of X.629
M1.b: Constant trend630
For each glacier with an incomplete time series, a linear trend is fitted to the available mass631
balance data and extrapolated to the unmeasured period [112]:632
bi(t) = a0,i + a1,i · t, (4)
where a0 and a1 are two coefficients to be estimated for glacier i, and t ∈ tn. The coefficients633
are estimated using least squares fit.634
M1.c: Constant ratio635
For each glacier with an incomplete time series, the ratio between its average mass balance636
and the average mass balance of the two continuously measured glaciers is computed. The637
annual mass balance during the unmeasured period is then computed by multiplying the average638
measured balance of the two measured glaciers with that ratio:639
bi(t) =1
2· (bj(t) + bk(t)) · bi(ty)
1/2 · (bj(ty) + bk(ty)), (5)
where j and k indicate the two glaciers for which a measured mass balance is available contin-640
uously between 1961 and 2012, and t ∈ tn.641
M1.d : Constant relation642
For each glacier with an incomplete time series, a relation to the mass balance of the two contin-643
uously measured glaciers is established through a multiple linear regression. The annual mass644
balance during the unmeasured period is then computed by using the continuously measured645
time series as input:646
bi(t) = c0,i + c1,i · bj(t) + c2,i · bk(t), (6)
where c0,i, c1,i, and c2,i, are the coefficients of the linear regression, j and k are the indices of647
the two glaciers with a continuously measured mass balance, and t ∈ tn. As for model M1.b,648
the coefficients are estimated using least squares fit.649
C.3 M2 : Estimates based on mass balance modelling650
Estimates of the group M2 are based on mass balance modelling. The glaciological in-situ mea-651
surements are used to calibrate two different models. The first one (M2.a) is based on a degree-652
day approach whilst the second (M2.b) relies on an energy-balance formulation. Both models are653
spatially distributed and are forced by meteorological reanalysis data. Additional local temper-654
ature and precipitation observations are used to constrain model parameters (see below). Three655
reanalysis products, ERA-40 [113], ERA-Interim [114] and NCEP/NCAR Reanalysis 1 [115], are656
considered in order to account for uncertainties in the meteorological input fields (Tab. C.3). The657
models are forced with the three time series individually, providing an ERA-40 driven run for658
the period 1961-2001, an ERA-Interim driven run for 1980-2012, and an NCEP/NCAR driven659
run for 1961-2012. In Table C.4, the three drivers are denoted with E4, EI, and NN, respec-660
tively. ERA-40 and ERA-Interim reanalysis data are retrieved from the portal of the European661
Centre for Medium-Range Weather Forecasts (http://apps.ecmwf.int/datasets/), whilst662
NCEP/NCAR data are provided by the Earth System Research Laboratory of the National663
Oceanic & Atmospheric Administration (NOAA, ftp.cdc.noaa.gov/Projects/Datasets/).664
Table C.3 gives an overview of the data used, whilst Figure C.1 illustrates the spatial distribu-665
tion and temporal evolution of the temperature and precipitation forcing fields for the addressed666
28
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 29: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/29.jpg)
region. Monthly values for individual sub-regions are shown in Figure C.2. Additional local tem-667
perature and precipitation data with daily resolution and a time series length of more than 5668
years are obtained for a total of 72 and 183 stations, respectively, from the Global Historical669
Climatology Network [116] maintained at NOAA’s National Climatic Data Center, the national670
hydro-meteorological services of Kyrgyzstan, Uzbekistan, and Tajikistan, and the Central-Asian671
Institute of Applied Geosciences. The location of the stations is displayed in Figure C.1.672
For any grid cell i within a given glacier, the models of group M2 compute daily mass balance673
bi (kg m−2 d−1) by subtracting daily ablation ci from accumulation ai, i.e. bi = ai − ci. The674
difference between the models M2.a and M2.b is given by the computation of ci, whilst ai is675
computed for both models as676
ai = Pref · (1 + Cprec) · [1 + (zi − zref) · dP/dz] ·Dsnow,i · rs, (7)
[117, 118]. In the equation, Pref (kg m−2 d−1) is the daily precipitation obtained from the reanal-677
ysis data and the reference altitude zref (see below); Cprec is a dimensionless calibration factor678
used to accommodate the local, measured annual accumulation; zi (m a.s.l.) is the elevation679
of the considered grid cell; dP/dz (m−1) is a precipitation lapse rate describing the relative680
increases in precipitation with altitude [119]; Dsnow,i is a dimensionless, spatially distributed681
factor which accounts for snow redistribution processes [e.g. 120, 121]; and rs (dimensionless)682
is the fraction of solid precipitation. The daily precipitation amount Pref is calculated through683
inverse-distance weighting of the precipitation given in the four reanalysis grid cells closest to684
the glacier centrepoint. The corresponding geopotentials of the four considered grid cells are685
weighted in the same way and define the reference altitude zref . Dsnow,i is determined from char-686
acteristics of the surface topography [117], dP/dz is estimated from the station data through687
linear regression of the annual precipitation sums and the station elevation, and rs is defined688
to decrease linearly from 1 to 0 in the temperature range Tthr − 1 ◦C to Tthr + 1 ◦C, where689
Tthr = 1.5 ◦C is a threshold temperature that distinguishes snow from rainfall [122].690
M2.a: Degree-day approach691
In the degree-day approach, ci is computed as692
ci =(fM + rsnow/ice · Ipot,i
)· T i if T i > 0 ◦C (8)
[122], where fM (kg m−2 d−1 ◦C−1) is a melt factor, rsnow/ice (kg W−1 d−1 ◦C−1) are two distinct693
radiation factors for snow and ice, Ipot,i (W m−2) is the potential direct clear-sky solar radiation694
for grid cell i, and T i (◦C) is the mean daily air temperature for the same grid cell. For days695
with T i ≤ 0 ◦C, no ablation occurs. The spatial distribution of T i is obtained by first computing696
a representative temperature for the reference elevation zref as described for precipitation, and697
then extrapolating the temperature over the glacier domain by means of a vertical temperature698
lapse rate dT/dz (◦C m−1). The lapse rate has a yearly cycle (one value for each day of the699
year) and is obtained through linear regression of the temperature observations at the local meteo700
stations. The spatial distribution of Ipot,i is first computed at hourly time steps according to the701
methods presented in [122], and then averaged to daily values. Thus, the averaging is performed702
over the temporal domain, whilst the spatial distribution is maintained.703
Following [123] for example, the parameters fM, rsnow, and rice are calibrated for each glacier704
individually by imposing a constant ratio of 0.015:0.75:1.0 between them, and minimizing the705
discrepancy between measured and observed annual balances. The assumption is required for706
reducing the degrees of freedom of the model. The ratio rsnow:rice of 0.75:1.0 is meant to reflect707
the difference between the broadband albedo of snow and ice, whilst the ratio fM:rice of 0.015:1.0708
was found to be appropriate in previous analyses [e.g. 124, 125, 118].709
Calibration is performed in an automated, iterative procedure. The model is initialized with710
an initial guess for Cprec and rice, and run forward in time. In a first step, Cprec is adjusted in711
29
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 30: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/30.jpg)
order to match the mean of the available annual accumulation measurements. In a second step,712
rice (and thus rsnow and fM) is adjusted in order to minimize the sum of the absolute deviations713
between modelled and measured annual mass balances. Since this re-introduces a mismatch714
in the accumulation measurements, Cprec is re-calibrated again, and so on. The procedure715
converges within 10 iteration steps in all cases.716
Through the chosen calibration procedure, the model parameters incorporate the effect of down-717
scaling the driving reanalysis data (which by design represent the larger-scale meteorological718
boundary conditions rather than the actual local meteorology) to the local conditions. Our719
model approach is, thus, an effective method for quantifying mass budgets, but does not allow720
us to study the full spectrum (e.g. mountain-induced mesoscale flow processes) of the large-721
to local-scale interactions. The latter would require a physically-based downscaling, which has722
only been established for individual glaciers so far [e.g. 126].723
In this respect some uncertainty remains in the calibrated values of Cprec since, in the region724
of interest, the highest rates of both accumulation and ablation occur during summer. In fact,725
although the available measurements are declared to be annual (i.e. total) accumulation values726
by the relevant data sources, some doubt remains on whether these measurements were not727
actually collected within a stratigraphic framework and thus rather represent net accumulation.728
In this latter case, total modelled accumulation would likely be underestimated. Note, however,729
than even in this case the presented results, which all refer to annual mass budgets only, would730
remain unaffected. This is because through the calibration to the annual mass balances, any731
underestimation of accumulation would be compensated by an overestimation of ablation, thus732
leaving the annual budgets unaltered.733
M2.b: Energy balance formulation734
In the energy balance model, the simplified formulation by [127] is adopted, according to which735
ci can be computed as:736 {ci = −ϕd,i/L if ϕd,i > 0
ϕd,i = τ(1− α) ·QE + C0 + C1 · T i, (9)
where ϕd,i (J m−2 d−1) is the daily mean surface energy flux at location i, L = 334 103 J kg−1737
is the latent heat of fusion of ice, τ (dimensionless) is the total atmospheric transmissivity, α738
(dimensionless) is the broadband surface albedo, QE (W m−2) is the extra-terrestrial irradiation,739
and C0 (W m−2) and C1 (W m−2 ◦C−1) are two empirical parameters to be calibrated. The740
term C0 +C1T i parametrizes the sum of the long-wave radiation balance and the turbulent heat741
exchange [127]. T i is obtained the same way as described for model M2.a, whilst the product742
τQE is approximated by scaling Ipot,i calculated for model M2.a with the ratio between actual743
and clear-sky surface solar radiation derived from the reanalysis data.744
Following [127], we (1) fix C1 at a value of 10 W m−2 ◦C−1, (2) use C0 as a calibration parameter745
for maximizing the agreement between calculated and observed annual mass balances, and (3)746
assign two different values to α depending on the presence of a snow cover. For snow and ice747
covered surfaces we use α = 0.7 and α = 0.4, respectively [e.g. 128]. Calibration is performed748
analogously to the degree-day approach.749
As for the estimates based on ICESat, debris-free and debris-covered surfaces are not treated750
separately (this is true for both model options M1.a and M2.b). This choice is motivated by751
the fact that (a) the majority of the glaciers in Tien Shan are clean-ice glaciers, i.e. not covered752
by debris (according to [129], the debris-covered area fraction for the central part of the Tien753
Shan is less than 5 %), (b) the comparison with geodetic mass balances derived for the selected754
sub-regions that include some large glaciers with a significant debris coverage [130, 129], does755
not indicates the models yielding too negative mass balances for these areas (see Section C.7 and756
Fig. 5a-d of the main article), as would be expected by postulating an insulating effect of the757
debris cover, and (c) a sensitivity experiment conducted with model M2.a.e1.NN for the region758
30
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 31: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/31.jpg)
analysed in [129] (i.e. the region of the Tien Shan including the largest debris-covered glaciers),759
shows that even the postulation of an overall melt-rate reduction by 75 % under debris covered760
areas (a very substantial reduction), would change the mass budget for the entire Tien Shan761
by 0.3 Gt a−1 only (not shown). The fact that debris-covered glaciers can lose mass similar to762
debris-free glaciers has, moreover, recently been highlighted by a series of studies focussing on763
the Himalayas [e.g. 131, 95, 84, 132].764
The mass balance time series resulting for the individual glaciers for which direct measurements765
are available are shown in Figure C.3.766
C.4 Regional extrapolation767
The approaches described above (M1.a-d and M2.a-b) only provide an estimate of the mass768
balance time series for the seven glaciers that have available long-term measurements. In order769
to extrapolate these results over the entire study region, three options (e1, e2, and e3) are770
explored: Option e1 simply assigns the average value to all remaining glaciers in the region;771
option e2 assigns a glacier specific value following the nearest neighbour principle; and option772
e3 uses an inverse distance interpolation. For the estimates of group M1, the extrapolated773
quantity is the annual glacier mass balance, whilst for group M2 we spatially extrapolate the774
model parameters and use the corresponding model for explicitly calculating a glacier-specific775
mass balance for every glacier in the region. More sophisticated extrapolation approaches,776
that could include an elevation dependency for example, were not considered in light of the777
small sample of glaciers that would be available for establishing such a relation. The choice is778
additionally motivated by the findings of [133] who concluded that “simple arithmetic averaging779
that completely neglects glacier characteristics is a robust alternative particularly if only few780
(< 10) mass balance series [...] are available”. The combination of the different models with781
the different extrapolation schemes and, for models of group M2, the different meteorological782
drivers, provides an ensemble with a total of 30 model options (Tab. C.4).783
For the methods of group M1, only the surface area is required, whilst for M2, a DEM of the784
surface of each individual glacier is derived by intersecting the according RGI outline with the785
ASTER GDEM version 2. In this case, the ASTER GDEM is preferred upon the SRTM DEM786
because of the higher spatial resolution, and because for the modelling, exact knowledge of the787
acquisition date is less important than for the ICESat derived estimates (cf. section B.1). In788
order to save computational time, the glacier specific DEMs extracted from the ASTER GDEM789
are resampled according to the glacier size. The resampling is chosen such that the computational790
domain remains below 4 × 104 grid cells for every glacier. This results in a resolution between791
30 and 190 m. The potential direct clear-sky solar radiation is calculated from the so obtained792
DEMs, as described in section C.3. Together with the aforementioned meteorological fields, this793
provides all necessary inputs for both models (M2.a and M2.b) in group M2.794
The modelling procedure as described so far yields so called “reference-surface balances” [134],795
i.e. mass balances that refer to a constant glacier hypsometry. The concept was introduced796
by [135] and has been suggested as to be “more useful for climate interpretation” [136, 137].797
When adopted in modelling studies, the concept has the convenient advantage that it does798
not require any update of the considered glacier surfaces. For hydrological applications and799
questions related to water resources management, however, the relevant quantity is given by the800
so called “conventional balance” (i.e. the mass balance that refers to the actual glacier geometry801
at any point in time), since it directly reflects the amount of water that is stored or released802
by the considered glaciers. In general, the link between conventional and reference-surface803
balances is not straightforward, since the effects of the evolving glacier surface can be complex.804
An analysis carried out by [138] over a sample of 36 Alpine glaciers and a period of about805
80 years, however, suggests that the difference between the two quantities can be reasonably806
approximated by a linear function. Here, the hypothesis of a linear relation is tested for the Ak-807
31
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 32: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/32.jpg)
Shiirak mountain range (region “A” in Fig. 1 of the main article), for which the data provided808
by [139] allow an assessment. To this end, all 12 models of group M2 that have a meteorological809
driver covering the period 1977-2000 (see Tab. C.4) are re-initialized three times: The first and810
second time, reference-surface balances are computed by using the surface DEMs of the year811
1977 and 2000 provided by [139], respectively; whilst the third time, the glacier surface of all812
the glaciers located in the mountain range is updated on a yearly basis. In this third case, the813
updating is performed by linearly interpolating the elevations of individual grid cells between814
the two years 1977 and 2000 [e.g. 140, 117]. The mass change rates resulting from the three815
model runs are shown in Figure C.4a and C.4b for the example of model option M2.a.e2.E4. The816
difference between the results obtained by computing conventional and reference-surface balances817
for the individual model options (Fig. C.4c), confirms that the magnitude of the difference818
grows approximately linearly with time to/from the reference year. This observation is used for819
converting the reference-surface balances computed for the entire study region into conventional820
mass balances: A linear function is fitted through ordinary least squares to the differences found821
in the case of the Ak-Shiirak region (solid line in Fig. C.4c), and the function is assumed to822
be applicable to all remaining areas of interest as well. Although the procedure has a clear823
advantage in terms of computational cost, the choice is mainly motivated by data availability.824
The considered glacier inventory and DEM (i.e. the RGI and the ASTER GDEM, respectively)825
are in fact the only data with a (nearly) time-consistent, region-wide coverage. Since both826
data sets refer to the end of the reconstructed period, a transient updating of the individual827
glacier surfaces would require a backward-in-time integration, which is not straightforward. The828
uncertainty introduced in the regional mass budget by converting the reference surface mass829
balances into conventional ones with the described procedure, is accounted for in the overall830
error budget (section C.6).831
C.5 Estimation of regional mass budgets832
For obtaining time series of mass budgets, the specific mass change rates calculated above need833
to be multiplied by the according glacier area at every time step. Since none of the models is834
able of updating the glacier area explicitly, glacier area changes are prescribed directly. This835
is done by compiling sub-regional area changes from the literature (Tab. C.2), and assuming a836
constant change rate during the period 1961-2012.837
It follows from simple geometrical considerations that for the same climatic forcing, small glaciers838
generally exhibit higher relative area changes than larger ones. Despite large variability, the839
results of various studies carried out in the region [e.g. 141, 142, 108] suggest that this size840
dependency can be roughly approximated through an empirical relation of the form841
r = a (1/A)b, (10)
where r is the relative area change of a glacier with area A, and a and b are two coefficients to be842
estimated. Here, the relation is used for distributing the total area change inferred for a given843
sub-region over the individual glaciers contained in the RGI. This is done by fixing parameter844
b in equation 10 to a constant value of 0.1 (the value is estimated from the results by [141],845
cf. their Figure 9) and adjusting parameter a until the total area change in the sub-region is846
matched. For the sample of glaciers that is not contained in any of the sub-regions for which an847
area change is available, the average rate of all other sub-regions is imposed.848
The resulting spatial distribution of the area changes for the period 1961-2012 is given in Fig-849
ure C.5. Numerical values for the area changes and additional information on the individual850
data sources are given in Table C.2. Resulting time series of actual mass change rates for the851
entire study region and the individual model options are shown in Figure 6 of the main article.852
For the period in which results derived from GRACE and ICESat data are available simulta-853
neously, the average mass change rate for the individual model options is reported in the last854
32
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 33: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/33.jpg)
column of Table C.4, and visualized in Figure 3c of the main article. The spatial distribution of855
the average mass change rate for the periods 1961-2012 and 2003-2009 are shown in Figures 1856
and 5e of the main article.857
C.6 Uncertainty estimates858
All confidence intervals stated for the estimates based on glaciological modelling are constructed859
according to the theory of Gaussian error propagation [e.g. 143]. All uncertainties are considered860
to be potentially correlated. In general, the variance var(f) of a function f with n variables861
x = (x1, x2, ..., xn)T, is given by862
var(f) = gTV g (11)
where g is the vector of size n whose ith element is ∂f∂xi
, and V is the variance-covariance matrix863
of x, with elements Vij = cov(xi, xj). In the notation, xi is the ith element of a vector x, xT864
denotes the transpose of x, ∂f∂xi
is the partial derivative of f with respect to xi, and cov(xi, xj)865
is the covariance between xi and xj . Note that cov(xi, xi) corresponds to the variance of xi,866
denoted with σ2xi
.867
When computing the total mass change ∆M of a region with n glaciers during a period of m868
years, the function f becomes:869
f = ∆M =
m∑t=1
n∑i=1
bi(t) ·Ai(t) (12)
where bi(t) is the estimated annual balance of glacier i for year t, and Ai(t) the corresponding870
glacier area. For this case, it can be shown that the evaluation of Equation 11 requires knowledge871
of (1) σ2bi(t)
, which is the variance of the estimated mass balance for glacier i and time t, (2)872
cov(bi(t), bj(t)), which is the covariance of the estimated mass balance of glaciers i and j at873
time t, (3) ρbAi(t), which is the correlation between estimated mass balance and glacier area874
for glacier i and time t, and (4) σ2Ai(t)
, which is the variance of the glacier area for glacier i875
and time t. For the analyses, σ2bi(t)
, cov(bi(t), bj(t)), and ρbAi(t) are estimated in a leave-one-876
out cross-validation scheme [144], and by using the seven measured mass balance time series877
available. For all 30 considered model options (Tab. C.4), the model set-up and calibration878
is performed seven times. At each time, six glaciers are used for calibration, whilst for the879
seventh, non-considered glacier, a mass balance time series is calculated and compared to the880
actual measurements. The so obtained differences are then used to estimate both an option-881
specific variance-covariance matrix, providing σ2bi(t)
and cov(bi(t), bj(t)) for the seven glaciers882
with measurements, and the correlation between estimated mass balance and glacier area. For883
the generalization to the whole region and supported by the analysis of the model residuals, we884
assume that time and glacier dependencies can be ignored, i.e. σ2bi(t)
= σ2b , with σ2
b being the885
average of σ2bi(t)
for the seven glaciers with measurements, and analogously for cov(bi, bj) and886
ρbAi. This is equivalent to the assumption that on average, the model performs equally well for887
any given glacier as it performs for a glacier with available measurements, when the latter is888
NOT used for calibration. For σ2Ai(t)
, which corresponds to the accuracy with which the area889
of glacier i is known at time t, it is conservatively assumed that the glacier area is known with890
a relative accuracy of ±50% at the 95 % confidence level at any given time. In the adopted891
notation, this means σ2Ai(t)
/A2i (t) = (0.5/2)2. The conservative estimate is chosen in order to892
accommodate the uncertainties deriving from both the RGI glacier outlines and the approach893
used for updating the glacier area (Section C.5).894
For the models of group M2, additional uncertainty is introduced by the procedure used for895
converting reference-surface balances to conventional ones. The uncertainty is included in the896
formulation used above by augmenting the standard error associated to the estimated mass bal-897
ance by σref.surf(t). The time dependency can not be neglected in this case, since the uncertainty898
33
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 34: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/34.jpg)
of the correction increases with increasing time distance from the reference year (cf. the size of899
the bars in Fig. C.4c). The results of the analysis carried out for the Ak-Shiirak mountain range900
indicate that the spread between the correction computed for different model options can be in901
the same order of magnitude of the correction signal itself. For any time t, σref.surf(t) is thus902
set to the magnitude of the applied correction.903
When presenting glacier changes as a fractional change in total glacier mass (the estimates for904
total mass is retrieved from [145]), the uncertainty in the latter needs to be accounted for. This905
is done by assuming (1) independence between estimated mass budgets and estimated total906
glacier mass, (2) stationarity in the uncertainty of the modelled mass budgets, and (3) a 10 %907
relative uncertainty (one-sigma level) in the estimated total glacier mass (cf. entry “Central908
Asia” in Table 2 of ref. [145]).909
C.7 Model validation910
The results of the different models of both groups M1 and M2 are validated against geodetic911
volume changes available from the literature. To date, only three regional assessments are912
available: [139] computed geodetic volume changes for the Ak-Shiirak mountain range (“Ak-913
Shiirak” in Fig. 1 of the main article; ∼370 km2 of glacierized area as of 2009) during the914
periods 1943-1977 and 1977-2000, [130] conducted a similar analysis for a selected sub-region915
south of Tomur Feng (Kyrgyz name: Jengish Chokusu; Russian name: Pik Pobedy) in the916
Aksu-Tarim catchment (“Tomur region” in Fig. 1 of the main article; ∼370 km2 of glaciers in917
2003) and the time periods 1976-1999 and 1999-2009, and this latter work was recently extended918
to the Aksu basin (“Aksu basin” in Fig. 1 of the main article; ∼5000 km2 of glaciers in 1999)919
and the period 1977-1999 by [129]. The comparison is not possible for every model option,920
since the meteorological forcing fields do not necessarily cover the entire time period (the period921
1943-1977, for example, is not covered entirely by any of the forcing fields). The results of922
the comparison are shown in Figure 5a-d of the main article. In general, the bulk of the model923
options is capable of reproducing the observed mass change rates. This is particularly true when924
considering the Aksu basin (Fig. 5a of the main article), which is the largest region for which925
a geodetic estimate of glacier mass changes is available. For the other regions, the picture is926
slightly more heterogeneous: For the sub-region south of Tomur Feng (“Tomur region” from now927
on) and the period 1999-2009, models of the group M2 reasonably agree with the results from the928
geodetic surveys, but one model M2.a.e2.EI (i.e. the degree-day model forced with ERA-Interim929
data and parameters assigned through a nearest-neighbour procedure) overestimates the mass930
loss rate. A similar overestimation is found for several models of the group M1, in particular931
for options relying on the temporal extrapolation of the observed trend (models of the category932
M1.b) or options based on a a nearest-neighbour procedure (models with ending .e3). This can933
be explained by the fact that for the period 1999-2009, the closest glacier with a measured time934
series (Tuyuksu Glacier, see Fig. 1 of the main article for location) lies about 300 km away from935
the Tomur region, and is thus not particularly representative. For the same region and the936
period 1976-1999, all models of the group M1 perform well. Models of the group M2, however,937
generally underestimate the mass change loss. The underestimation is very prominent in model938
runs driven by ERA-40 data, and can be attributed to a significant precipitation anomaly in939
the dataset for the particular region and the considered period. This points at the fact that940
the results of individual model options must be interpreted with caution, especially at the sub-941
regional scale. The mass change rates of the Ak-Shiirak mountain range is well reproduced by942
models of the group M1. Although consistent within the assessed confidence intervals, models943
of the group M2 indicate a less negative mass change rate than M1. This difference reflects944
the fact that Sary-Tor Glacier (the closest glacier with in-situ measurements, see Fig. 1 of the945
main article) is located within the considered mountain range, but has an altitude that is below946
average compared to the region. This results in a generally more negative mass balance for the947
34
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 35: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/35.jpg)
particular glacier, and has a large effect in the models of group M1 which rely on the measured948
time series only. On the other hand, models of group M2 are capable of accounting for this949
elevation difference and thus result in a less negative mass change rate. The fact that even950
models of group M1 yield a mass change rate that is less negative than reported by [139] seems951
to suggest that the assessed value might have been too negative.952
The obtained spatial distribution of the modelled mass budgets is additionally compared to sub-953
regional estimates derived from ICESat. To this end, the analyses described in Section B are954
repeated individually for each sub-region defined in Figure 5e of the main article. The number of955
ICESat footprints available within the individual sub-regions is limited, which is reflected in the956
comparatively large confidence intervals of the derived estimates (Table C.5). For sub-regions in957
which a sufficient number of footprints (about 2000) are available, the agreement is satisfactory958
(Fig. 5f of the main article). This supports the capability of the model ensemble in capturing the959
large-scale spatial variability of the mass change signal. An exception seems to be the Borohoro960
range (region “R3”) for which the ICESat estimates indicate a more negative mass change than961
the glaciological models do. A partial explanation could be the relative under-sampling of very962
high elevations in the ICESat footprints for the particular region (not shown).963
C.8 Multiple linear regression analysis964
In general, a multiple linear regression (MLR) model aims at reproducing the n values observed965
for a target variable y through a linear combination of a series of m explanatory variables X,966
i.e.967
y = Xβ + ε. (13)
In the notation, y = (y1, y2, ..., yn)T is the vector of n values observed for the target variable;968
X is a (n × (m + 1)) matrix with elements Xi1 = 1 for i ∈ [1, n] and elements Xi(j+1) = x(j)i969
for i ∈ [1, n], j ∈ [1,m], with x(j)i being the ith observed value for the jth explanatory variable;970
β = (β0, β1, β2, ..., βm)T is a set of m+ 1 coefficients to be estimated; and ε = (ε1, ε2, ..., εn)T is971
a vector of residuals for which a normal distribution with zero mean and standard deviation σ972
is assumed. The estimation of β is performed through ordinary least squares fit, i.e. through973
the minimization of∑
y −Xβ, where β is the vector of estimated coefficients. The fraction of974
variance in the target variable that is explained by the combination of a given set of explanatory975
variables, can be expressed through the adjusted coefficient of determination R2 [e.g. 146].976
For detecting which topographical parameters have the largest influence on the results provided977
by the models of group M2, a first MLR model that includes all possible explanatory variables978
is set up. This first set of explanatory variables includes longitude, latitude, elevation, elevation979
range, glacier area, slope, and orientation, calculated as the mean value for every individual980
glacier. The target variable is defined as the average glacier-specific mass change rate over the981
period 1961-2012. A common backwards selection procedure [e.g. 147], i.e. a one-at-the-time982
removal of non-significant variables, is not suitable for model reduction since the large sample983
of glaciers (about 13700) yields significant coefficients for all variables (all p-values < 10−16).984
An “all possible subsets regression procedure” [e.g. 148] is therefore preferred. To this end, all985
26 = 64 possible MLR models that can be obtained from the combinations of the 6 explanatory986
variables are set up, and evaluated in terms of R2. Plotting this set of R2 coefficients in a Pareto987
chart [e.g. 149] reveals a clear break-point when elevation, orientation, and latitude are included988
in the MLR model (Fig. C.6a). Together, the three variables explain about 75% of the total989
variance of the target variable. Elevation alone explains 41% of the variance.990
The analogous procedure is repeated for detecting the variables having the strongest influence in991
the temporal variability. To this end, standardized yearly anomalies in temperature, precipita-992
tion, downscaled net solar radiation, and positive degree days are computed for each individual993
glacier, and averaged over the entire mountain range. The standardized annual glacier mass994
35
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 36: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/36.jpg)
change anomaly is chosen as target variable. In this case, the break point in the Pareto chart995
is found for MLR models that include precipitation and positive degree days (Fig. C.6b). Com-996
bined, the two variables explain 93% of the total variance. The sum of positive degree days997
alone explains 83% of the variance.998
36
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 37: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/37.jpg)
ER
A-4
0
-20 0 20
Temperature ( oC)
ER
A-I
nte
rim
NC
AR
/NC
EP
72o E 74o E 76o E 78o E 80o E 82o E 84o E 86o E
40o N
42o N
44o N
0 1000 2000
Precipitation ( mm a-1 )
40o N
42o N
44o N
72o E 74o E 76o E 78o E 80o E 82o E 84o E 86o E
40o N
42o N
44o N
1960 1970 1980 1990 2000 2010
-2
-1
0
1
2
An
om
aly
w.r
.t. r
efer
ence
reference
1960 1970 1980 1990 2000 2010
-200
-100
0
100
200
300
referenceERA-40
ERA-Interim
NCAR/NCEP
Figure C.1: Overview of the used meteorological forcing fields. The left column refers to air temperature,the right column to precipitation. Starting from the top, the first three panels of each column show thespatial distribution of the mean value during the reference period 1979-2001 at the resolution of theparticular reanalysis product (either ERA-40, ERA-Interim, or NCEP/NCAR). The reference period ischosen as the period in which all three reanalysis data sets have values. The lowermost panel in eachcolumn is a time series of yearly anomalies with respect to the reference period. In this case, values referto the average of the displayed domain. The yellow (red) dots in the uppermost three panels indicate thelocation of the available temperature (precipitation) stations.
37
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 38: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/38.jpg)
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
455 mm
-10
0
10
20 Temperature (
oC)
-0.6 oC
J F M A M J J A S O N D
R1
0
400
800
Are
a (k
m2 ) 5194 km2
4428 m
N
50%
25
Aspect
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
403 mm
-10
0
10
20 Temperature (
oC)
-0.7 oC
J F M A M J J A S O N D
R2
0
400
800
Are
a (k
m2 ) 2007 km2
3920 m
N
50%
25
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
373 mm
-10
0
10
20 Temperature (
oC)
1.5 oC R3
0
400
800
Are
a (k
m2 ) 2734 km2
3938 m
N
50%
25
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
547 mm
-10
0
10
20 Temperature (
oC)
-0.3 oC R3a
0
400
800
Are
a (k
m2 ) 1474 km2
4084 m
N
50%
25
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
491 mm
-10
0
10
20 Temperature (
oC)
4.1 oC R4
0
400
800
Are
a (k
m2 ) 649 km2
3600 m
N
50%
25
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
483 mm
-10
0
10
20 Temperature (
oC)
2.9 oC R5
0
400
800
Are
a (k
m2 ) 640 km2
3966 m
N
50%
2550
100
150
200
Pre
cipi
tatio
n (m
m)
0
516 mm
-10
0
10
20 Temperature (
oC)
-1.3 oC R6
3000 4000 5000 6000Elevation (m a.s.l.)
0
400
800
Are
a (k
m2 ) 1768 km2
4185 m
N
50%
25
50
100
150
200
Pre
cipi
tatio
n (m
m)
0
620 mm (B)
-10
0
10
20 Temperature (
oC)
1.9 oC (A) All other
3000 4000 5000 6000Elevation (m a.s.l.)
0
400
800
Are
a (k
m2 ) 718 km2
3883 m (C)(D)
N
50%
25
Aspect
Aspect
Aspect
Aspect
Aspect
Aspect
Aspect
Figure C.2: Seasonal distribution of the meteorological forcing fields and topographic characteristics forindividual sub-regions. For each sub-region “R1” to “R6” (see Fig. 5e of the main article for location)three panels are displayed. The top panel shows monthly means for air temperature (solid red line) andprecipitation (blue bars). Mean values are obtained by averaging over the three reanalysis products (ERA-40, ERA-Interim, NCEP/NCAR) and the reference period 1979-2001. Vertical confidence intervals showthe range spanned by the individual reanalysis products. The bottom left panel shows the hypsometry(i.e. the distribution of area with elevation) of the glacierized surfaces. The hypsometry is obtained byconsidering mean glacier elevations and 100 m elevation bins. The radar chart in the bottom right panelshows the aspect of the glacierized surfaces. The aspect is expressed as the share of the total glacierizedarea, and is displayed for the eight sectors N, NE, E, SE, S, SW, W, and NW. The last set of panels(labelled with “All others”) refers to glacierized surfaces that are not included in any of the sub-regionsshown previously. The set includes the key for the numbers displayed in the panel corners: (A) = Averageannual temperature at mean glacier elevation; (B) = Mean annual precipitation; (C) = Total glacierizedarea; (D) = Mean glacier elevation.
38
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 39: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/39.jpg)
1960 1970 1980 1990 2000 2010
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Urumqi Gl. No.1 (URU)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Shumskiy (SHM)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Ts. Tuyuksu (TYK)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Golubin (GLB)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Sary-Tor (SRT)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Kara-Batkak (KRB)
-3
-2
-1
0
1
MB
(m w
.e. a
-1 )
Abramov (ABR)
MeasuredModelled
Figure C.3: Mass balance time series for glaciers with available measurements. For each glacier, theobserved (black) and modelled (red) mass balance (MB) is shown. The modelled mass balance is given asthe mean of all model options listed in Table C.4. The grey band is the envelope of all model realizationsof group M2. The location of the individual glaciers is shown in Figure 1 of the main article. The key forthe three-letter code on the top right corner is given in Table C.1. For SRT, the solid grey line indicatesthe reconstruction by [150].
39
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 40: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/40.jpg)
1980 1985 1990 1995 2000year
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
Gla
cier
mas
s ch
ange
rat
e(1
03 kg
m-2 a
-1 )
ConventionalRef. surface 1977Ref. surface 2000
Ak-Shiirak
Option M2.a.e2.E4
a
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
Ave
rage
mas
s ch
ange
rat
e(1
03 kg
m-2 a
-1 )
conv
.r.1
977
r.200
0
b
-30 -20 -10 0 10 20 30Time from reference year (a)
-0.1
0.0
0.1
Diff
eren
ce (
103 k
g m
-2 a
-1 )
conv
entio
nal v
s re
f. surf
ace
All available options c
Figure C.4: Difference between mass change rates computed by considering conventional and reference-surface mass balances. (a) Time series of the specific mass change rate dm/dt computed for the glaciersin the Ak-Shiirak mountain range (see Fig. 1 of the main article) during the period 1977-2000. The black(blue and red) line shows the dm/dt resulting from the computation of conventional (reference-surface)mass balances. (b) Mass change rates averaged over the considered region and the considered periodaccording to the three methods. Confidence intervals refer to the 95% level. (c) Difference between masschange rates computed according to the conventional and the reference-surface method as function of thetime from the reference year. Blue (red) bars refer to the case in which the reference-surface correspondsto the year 1977 (2000), and span over the range given by the 12 considered model options (the minimalspan is fixed to 2.5 103 kg m−2 a−1 for visibility). The solid black line is a linear function fitted to all datapoints through ordinary least squares. Panels (a) and (b) refer to the model option M2.a.e2.E4, whilstpanel (c) comprehends all 12 options for which meteorological drivers are available for the consideredtime period (see Tab. C.4 for details).
40
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 41: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/41.jpg)
200 kmSurface elevation (m a.s.l.)
0 3000 6000
72o E 74o E 76o E 78o E 80o E 82o E 84o E 86o E
40o N
42o N
44o N
A01
A02b
A02a
A02c
A07a
A03
A02e
A02dA04
A05
A06A07
A08
A09
A10
A11
A12aA12bA17 A16
A14
A12dA12c
A15
A13
Glacier area change 1961-2012 (%)
0 -10 -20 -30 -40 -50
Figure C.5: Spatially distributed glacier area changes for the period 1961-2012, relative to the glacierarea in 1961. Each coloured dot represents one glacier within the RGI v3.2. Data sources and detailedinformation for individual sub-regions are given in Table C.2. For additional details on the symbology,refer to Figure 1 of the main article.
0.0
0.2
0.4
0.6
0.8
1.0
Por
tion
of e
xpla
ined
var
ianc
e
elevationaspectlatitude
areaslope
longitude
000000
001000
000001
000010
001010
000100
000011
001100
001011
001101
001001
000101
000110
001110
010000
011000
000111
001111
010001
011001
010010
011010
011011
010011
010100
011100
010101
011101
010110
011110
011111
010111
100010
100000
100001
100011
100100
100110
100101
100111
110000
110010
101000
101010
110100
110110
110001
101100
101110
110011
110101
110111
101001
101011
101101
101111
111000
111100
111010
111001
111110
111011
111101
111111
rad.110010100110101
a b
Included variables
elevationaspect or latitude
aspect and latitude
temp.000111110010011prec.011001100001111PDD000000011111111
Figure C.6: Pareto charts for the detection of variables strongly influencing the results. The portionof explained variance is shown for all possible regression models that relate (a) topographic variablesto the long-term (1961-2012) specific mass change rate of each glacier, and (b) meteorological variablesto the standardized annual mass budget anomaly of the region. The lower part of the figure indicateswhether a particular variable is included (1) or not included (0) in a particular regression model. In (b),the variables are standardized yearly anomalies of positive degree days (PDD), precipitation sum (prec.),average temperature (temp.), and net solar radiation (rad.).
41
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 42: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/42.jpg)
MonthJ F M A M J J A S O N D
Pre
ssur
e (h
Pa) 200
300
500-0.2
-0.1
0
0.1
0.2
0.3
0.4
Cor
rela
tion
coef
ficie
nt
Figure C.7: Linkage between glacier mass budgets and atmospheric circulation. The plot displaysthe correlation between the annual regional glacier mass-budget (ensemble mean of all model optionsin Table C.4) and the monthly meridional wind speed over the Tien Shan (averaged over the domain72◦-86◦ E, 40◦-45◦ N, data from NCEP/NCAR Reanalysis 1) at different atmospheric pressure levels.The analysis refers to the period 1961-2012. Both variables are linearly detrended. Black dots indicatesignificance at the 95 % confidence level.
42
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 43: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/43.jpg)
Table C.1: Overview of glaciological mass balance measurements used in this study, and period of dataavailability. For glaciers, TLC is a three letter code used for abbreviation. Location coordinates, given indecimal degrees, are approximate and refer to the centrepoint of the according glacier or region. “Meth.”indicates the method of data collection: stake readings (st. r.) or geodetic mass balance (geod.). “Ref.”is a reference for the data source. Individual glaciers are sorted according to glacier area.
TLC Glacier/Region Name Location Meth. Period Ref.
ABR Abramov 39.623 N 71.557 E st. r. 1968-1998 [151]KRB Kara-Batkak 42.100 N 78.300 E st. r. 1957-1998 [151]GLB Golubin 42.452 N 74.497 E st. r. 1969-1994 [151]SRT Sary-Tor 41.825 N 78.177 E st. r. 1985-19891 [151]TYK Ts. Tuyuksu 43.045 N 77.079 E st. r. 1957-2010 [151]SHM Shumskiy 45.083 N 80.233 E st. r. 1967-1991 [152]URU Urumqi Gl. No.1 43.116 N 86.809 E st. r. 1959-2010 [151]
– Ak-Shiirak 41.850 N 78.330 E geod. 1943-20002 [139]– Tomur region 41.900 N 80.100 E geod. 1976-20093 [130]– Aksu basin4 41.900 N 80.100 E geod. 1977-1999 [129]
1 A reconstructed glacier mass balance for the period 1930-1988 is available from [150]
and was included in the model calibration.2 [139] report two separate values for the periods 1943-1977 and 1977-2000, respectively.3 [130] report two separate values for the periods 1976-1999 and 1999-2009, respectively.4 The region is named Central Tien Shan in the original publication [129]
43
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 44: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/44.jpg)
Tab
leC
.2:
Ove
rvie
wof
the
dat
au
sed
for
qu
anti
fyin
ggl
acie
rar
each
ange
s.T
he
ind
ivid
ual
sub
-reg
ion
sar
ed
efin
edin
Fig
ure
C.5
.nRGI,A
RGI
=to
tal
gla
cier
nu
mb
eran
dar
ea,
resp
ecti
vely
,acc
ord
ing
toth
eR
GI
v3.
2;ndA
=nu
mb
erof
anal
yse
dgl
acie
rs;A
yr1
,A
yr2
,A
yr3
=gl
acie
rize
dar
eafo
rth
eyea
rsst
ate
din
the
colu
mn
syr 1
,yr 2
,yr 3
,re
spec
tive
ly;dA/dt
=an
nu
alp
erce
ntu
algl
acie
rar
each
ange
refe
ren
ced
toth
eye
aryr 1
;dA
ref
=p
erce
ntu
algl
acie
rar
each
an
ge
du
rin
gth
ep
erio
d19
61-2
012,
calc
ula
ted
asd
escr
ibed
inse
ctio
nC
.5.
Ifn
otst
ated
oth
erw
ise
abov
e,th
enu
mb
ers
are
retr
ieve
dfr
omth
eso
urc
egi
ven
inth
eco
lum
nR
ef..n.a.
ind
icat
esth
at
the
info
rmat
ion
isn
otav
ailab
le.
Con
fid
ence
inte
rval
sst
ated
for
the
resu
lts
ofth
isst
ud
yre
fer
toth
e95
%co
nfi
den
cele
vel.
Note
that
the
spel
lin
gu
sed
for
the
nam
esof
the
ind
ivid
ual
regi
ons
foll
ows
the
dat
aso
urc
egi
ven
inth
eco
lum
nR
ef.
(t.s
.st
and
sfo
r“t
his
stu
dy”)
,an
dm
ight
diff
erfr
omw
hat
state
din
Fig
ure
5eof
the
main
arti
cle
an
dT
able
C.5
.
Sub-r
egio
nnRGI
ARGI
ndA
yr 1
yr 2
yr 3
Ayr1
Ayr2
Ayr3
dA/dt
dA
ref
Ref
.-
km
2-
--
-km
2km
2km
2%
a−1
%
A01
Pask
emare
a28
34.0
525
1968
2000
2007
219.8
177.0
168.7
n.a
.-2
9.4
[141]
A02a
Low
erN
ary
nare
a199
88.4
n.a
.1977
1999
2007
83.0
77.0
75.0
n.a
.-1
5.7
[107]
A02b
At-
Bash
iK
irka
siare
a345
214.0
n.a
.1977
1999
2007
151.0
130.0
128.0
n.a
.-2
5.2
[107]
A02c
Bork
old
oyT
oo
are
a433
185.9
n.a
.1977
1999
2007
234.0
218.0
190.0
n.a
.-2
5.9
[107]
A02d
Big
Nary
nbasi
n626
468.7
462
1956
2007
n.a
.403.6
289.4
n.a
.n.a
.-2
9.1
[153]
A02e
Dzh
etim
are
a629
384.7
n.a
.1977
1999
2007
532.0
410.0
351.0
n.a
.-4
8.1
[107]
A03
SE
-Fer
gana
are
a119
94.6
306
1968
2001
2007
190.1
172.6
171.7
n.a
.-1
3.0
[141]
A04
At-
Bash
yare
a199
122.6
192
1968
2000
2007
113.6
99.9
95.7
n.a
.-1
9.6
[141]
A05
Aksu
Riv
erbasi
n907
1611.4
247
1963
1999
n.a
.1760.7
1702.1
n.a
.-0
.16
-8.2
[154]
A06
Ogan
Riv
erbasi
n847
1771.5
n.a
.1970
2002
n.a
.n.a
.n.a
.n.a
.-0
.31
-15.8
[142]
A07
Kaid
uR
iver
basi
n853
576.6
462
1963
2000
n.a
.331.1
292.6
n.a
.-0
.31
-15.9
[154]
A07a
Alb
inm
ounta
ins
71
66.4
70
1963
2000
n.a
.55.0
48.0
n.a
.n.a
.-1
7.4
[155]
A08
Aydin
gkol
Lake
basi
n79
40.1
203
1962
2006
n.a
.144.1
112.9
n.a
.-0
.56
-25.0
[156]
A09
Manas
Riv
erbasi
n1689
1427.0
n.a
.1962
1993
n.a
.n.a
.n.a
.n.a
.-0
.54
-27.5
[142]
A10
Ebin
ur
Lake
basi
n967
865.4
446
1964
2004
n.a
.366.3
312.5
n.a
.-0
.38
-18.5
[157]
A11
Ili
Riv
erbasi
n2002
1968.8
n.a
.1960
2009
n.a
.n.a
.n.a
.n.a
.-0
.46
-23.5
[142]
A12a
Sary
-Jaz
East
ern
regio
n390
809.5
318
1990
2010
n.a
.926.8
912.8
n.a
.n.a
.-3
.8[1
08]
A12b
Sary
-Jaz
Nort
her
nre
gio
n468
522.3
384
1990
2010
n.a
.487.4
455.8
n.a
.n.a
.-1
5.1
[108]
A12c
Sary
-Jaz
Wes
tern
regio
n478
443.4
498
1990
2010
n.a
.510.7
485.2
n.a
.n.a
.-1
1.9
[108]
A12d
Sary
-Jaz
South
ern
regio
n234
242.7
146
1990
2010
n.a
.130.1
124.1
n.a
.n.a
.-1
1.0
[108]
A13
Ak-S
hiira
kare
a222
302.2
178
1943
1977
2003
424.7
406.8
371.6
n.a
.-1
0.7
[139]
A14
Tes
key
are
a357
298.3
269
1971
2002
n.a
.245.0
226.0
n.a
.n.a
.-1
2.4
[158]
A15
Ili-
Kungoy
are
a831
616.9
735
1972
2000
2007
672.2
590.3
564.2
n.a
.-2
2.0
[141]
A16
Ala
-Arc
ha
are
a68
41.1
48
1963
1981
2003
42.8
40.6
36.3
n.a
.-1
9.3
[139]
A17
Sokolu
kR
iver
basi
n26
61.6
77
1963
1986
2000
31.7
27.5
22.8
n.a
.-3
6.8
[159]
–A
ll”re
main
ing”
gla
cier
s1809
1370.8
1809
1961
2001
2012
1690±
90
1370±
70
1300±
70
-0.4
5±
0.1
2-2
3.1±
6.6
t.s.
Tien
Shan
mounta
inra
nge
13696
13714.4
13696
1961
2001
2012
16150±
750
13710±
690
13190±
680
-0.3
6±
0.1
2-1
8.3±
6.4
t.s.
44
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 45: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/45.jpg)
Table C.3: Overview of the used reanalysis data. The name of the variable is stated as listed in theplatform from which the data were retrieved.
Data set Resolution Variable
ERA-40 1◦ × 1◦ 2 m air temperatureERA-40 1◦ × 1◦ Large-scale precipitationERA-40 1◦ × 1◦ Convective precipitationERA-40 1◦ × 1◦ Surface net solar radiationERA-40 1◦ × 1◦ Surface net solar radiation, clear sky
ERA-Interim 0.5◦ × 0.5◦ 2 m air temperatureERA-Interim 0.5◦ × 0.5◦ Total precipitationERA-Interim 1◦ × 1◦ Surface net solar radiationERA-Interim 1◦ × 1◦ Surface net solar radiation, clear sky
NCEP/NCAR 1.9◦ × 1.9◦ Mean daily air temperature at 2 mNCEP/NCAR 1.9◦ × 1.9◦ Mean daily precipitation rate at surfaceNCEP/NCAR 1.9◦ × 1.9◦ Mean daily clear sky downward solar flux at surfaceNCEP/NCAR 1.9◦ × 1.9◦ Mean daily downward solar radiation flux at surface
45
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 46: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/46.jpg)
Table C.4: Overview of the 30 different model options for estimates based on glaciological observationsand modelling. Glacio. meas. indicates that the model is driven by glaciological observations only.Other model drivers are defined in Table C.3. dM/dt (Gt a−1) is the mass change rate during the period2003-2009 in which both GRACE and ICESat data are available simultaneously. N.A. indicates that themodel driver is not available for the whole period. Confidence intervals refer to the 95% level. For theaverage, the stated confidence interval corresponds to two times the standard deviation of all ensemblemembers.
Code Group Model Extrapolation Driver Period dM/dt
M1.a.e1 Average Glacio. meas. 1961-2012 −4.91 ± 3.36M1.a.e2 Constant rate Nearest neighbour Glacio. meas. 1961-2012 −5.19 ± 3.44M1.a.e3 Inverse distance Glacio. meas. 1961-2012 −4.59 ± 3.34
M1.b.e1 Average Glacio. meas. 1961-2012 −8.27 ± 3.37M1.b.e2 Constant trend Nearest neighbour Glacio. meas. 1961-2012 −8.93 ± 4.13M1.b.e3 Inverse distance Glacio. meas. 1961-2012 −8.33 ± 3.37
M1M1.c.e1 Average Glacio. meas. 1961-2010 −6.54 ± 2.66M1.c.e2 Constant ratio Nearest neighbour Glacio. meas. 1961-2010 −7.47 ± 3.16M1.c.e3 Inverse distance Glacio. meas. 1961-2010 −6.29 ± 2.65
M1.d.e1 Average Glacio. meas. 1961-2010 −4.51 ± 2.55M1.d.e2 Const. relation Nearest neighbour Glacio. meas. 1961-2010 −5.34 ± 2.85M1.d.e3 Inverse distance Glacio. meas. 1961-2010 −4.45 ± 2.54
M2.a.e1.E4 ERA-40 1961-2001 N.A.M2.a.e1.EI Average ERA-Interim 1979-2012 −7.23 ± 2.56M2.a.e1.NN NCEP/NCAR 1961-2012 −7.31 ± 2.62
M2.a.e2.E4 ERA-40 1961-2001 N.A.M2.a.e2.EI Degree-day Nearest neighbour ERA-Interim 1979-2012 −8.17 ± 3.42M2.a.e2.NN NCEP/NCAR 1961-2012 −6.81 ± 2.55
M2.a.e3.E4 ERA-40 1961-2001 N.A.M2.a.e3.EI Inverse distance ERA-Interim 1979-2012 −5.53 ± 2.53M2.a.e3.NN NCEP/NCAR 1961-2012 −5.66 ± 2.47
M2M2.b.e1.E4 ERA-40 1961-2001 N.A.M2.b.e1.EI Average ERA-Interim 1979-2012 −5.41 ± 1.94M2.b.e1.NN NCEP/NCAR 1961-2012 −6.71 ± 2.51
M2.b.e2.E4 ERA-40 1961-2001 N.A.M2.b.e2.EI Energy balance Nearest neighbour ERA-Interim 1979-2012 −5.08 ± 2.47M2.b.e2.NN NCEP/NCAR 1961-2012 −6.28 ± 2.47
M2.b.e3.E4 ERA-40 1961-2001 N.A.M2.b.e3.EI Inverse distance ERA-Interim 1979-2012 −3.77 ± 2.09M2.b.e3.NN NCEP/NCAR 1961-2012 −5.08 ± 2.39
AVERAGE 2003-2009 −6.16 ± 2.84
46
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 47: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/47.jpg)
Table C.5: Average mass change rates for sub-regions defined in Figure 5 of the main article. dm/dtICESat
and dm/dtmodelling are the average specific mass change rate calculated from ICESat data and the modeloptions listed in Table C.4, respectively. The superscript indicate the period to which the estimaterefers to. The number of glaciers (ngl), the total glacier area (ARGI), and the number of retrieved ICESatfootprints over glaciers (nICESat) is given for each sub-region. ngl and ARGI are according to the RGI v3.2.Confidence intervals refer to the 95% level. The ICESat estimate for the remaining areas is not available(N.A.) since too few ICESat footprints can be retrieved. The estimates for the period 2003-2009 areplotted against each other in Figure 5 of the main article.
ngl ARGI dm/dt1961−2012modelling dm/dt2003−2009
modelling nICESat dm/dt2003−2009ICESat
Sub-region- km2 103 kg m−2 a−1 103 kg m−2 a−1 - 103 kg m−2 a−1
R1 Central Tien Shan 3087 5197.8 −0.21 ± 0.26 −0.31 ± 0.32 7157 −0.06 ± 0.31R2 Halik Shan 2429 2005.8 −0.63 ± 0.31 −0.69 ± 0.28 2945 −0.68 ± 0.43R3 Borohoro 3208 2734.9 −0.32 ± 0.30 −0.41 ± 0.28 2941 −0.63 ± 0.50R3a Central Borohoro 1449 1474.7 −0.17 ± 0.24 −0.29 ± 0.24 1432 −0.46 ± 0.76R4 Djungar Alatau 984 649.3 −0.44 ± 0.24 −0.49 ± 0.22 628 −0.75 ± 0.52R5 Ile and Kungoy Alatau 876 640.1 −0.31 ± 0.16 −0.33 ± 0.16 887 −0.68 ± 0.44R6 Inner Ranges 2303 1768.9 −0.36 ± 0.21 −0.41 ± 0.22 2336 −0.39 ± 0.37– Remaining areas 812 718.4 −0.60 ± 0.47 −0.72 ± 0.49 248 N.A.
47
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 48: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/48.jpg)
Supplementary References999
[1] B.D. Tapley, S. Bettadpur, M.M Watkins, and C. Reigber. The gravity recovery and1000
climate experiment; mission overview and early results. Geophysical Research Letters,1001
31(9):L09607, 2004. doi: 10.1029/2004GL019920.1002
[2] A. Cazenave and J. Chen. Time-variable gravity from space and present-day mass redis-1003
tribution in the Earth system. Earth and Planetary Science Letters, 298:263–274, 2010.1004
doi: 10.1016/j.epsl.2010.07.035.1005
[3] G. Ramillien, J.S. Famiglietti, and J. Wahr. Detection of continental hydrology and1006
glaciology signals from GRACE: a review. Surveys in Geophysics, 29(4-5):361–374, 2008.1007
doi: 10.1007/s10712-008-9048-9.1008
[4] C. Jekeli. The determination of gravitational potential differences from satellite-to-1009
satellite tracking. Celestial Mechanics and Dynamical Astronomy, 75(2):88–101, 1999.1010
doi: 10.1023/A:1008313405488.1011
[5] D.D. Rowlands and S.B. Luthcke. Resolving mass flux at high spa-1012
tial and temporal resolution using GRACE intersatellite measurements.1013
http://onlinelibrary.wiley.com/doi/10.1029/2004GL021908/full, 32(4):L04310, 2005.1014
doi: 10.1029/2004GL021908.1015
[6] K. Lambeck. Geophysical geodesy: The study of the slow deformations of the Earth, pages1016
7–10. American Geophysical Union, 2013. doi: 10.1029/GM060p0007.1017
[7] R. Klees, X. Liu, T. Wittwer, B.C. Gunter, E.A. Revtova, R. Tenzer, P. Ditmar, H.C.1018
Winsemius, and H.H.G. Savenije. A comparison of global and regional GRACE models1019
for land hydrology. Surveys in Geophysics, 29(4-5):335–359, 2008. doi: 10.1007/s10712-1020
008-9049-8.1021
[8] D.D. Rowlands, S.B. Luthcke, J.J. McCarthy, S.M. Klosko, D.S. Chinn, F.G. Lemoine,1022
J.-P. Boy, and T.J. Sabaka. Global mass flux solutions from GRACE: A comparison of1023
parameter estimation strategiesMass concentrations versus Stokes coefficients. Journal of1024
Geophysical Research: Solid Earth, 115:B01403, 2010. doi: 10.1029/2009JB006546.1025
[9] J. Wahr, M. Molenaar, and F. Bryan. Time variability of the Earth’s gravity field: Hy-1026
drological and oceanic effects and their possible detection using GRACE. Journal of Geo-1027
physical Research: Solid Earth, 103(B12):30205–30229, 1998. doi: 10.1029/98JB02844.1028
[10] J.Y. Guo, Y.B. Li, Y. Huang, H.T. Deng, S.Q. Xu, and J.S. Ning. Green’s function of1029
the deformation of the Earth as a result of atmospheric loading. Geophysical Journal1030
International, 159(1):53–68, 2004. doi: 10.1111/j.1365-246X.2004.02410.x.1031
[11] F. J. Simons and F.A. Dahlen. Spherical Slepian functions and the polar gap in1032
geodesy. Geophysical Journal International, 166(3):1039–1061, 2006. doi: 10.1111/j.1365-1033
246X.2006.03065.x.1034
[12] R. Klees, E.A. Zapreeva, H.C. Winsemius, and H.H.G. Savenije. The bias in GRACE1035
estimates of continental water storage variations. Hydrology and Earth System Sciences,1036
11(4):1227–1241, 2007. doi: 10.5194/hess-11-1227-2007.1037
[13] S. Swenson and J. Wahr. Methods for inferring regional surfacemass anomalies1038
from Gravity Recovery and Climate Experiment (GRACE) measurements of timevari-1039
able gravity. Journal of Geophysical Research: Solid Earth, 107:1–13, 2002. doi:1040
10.1029/2001JB000576.1041
48
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 49: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/49.jpg)
[14] L. Longuevergne, B.R. Scanlon, and C.R. Wilson. GRACE Hydrological estimates for1042
small basins: Evaluating processing approaches on the High Plains Aquifer, USA. Water1043
Resources Research, 46(11):W11517, 2010. doi: 10.1029/2009WR008564.1044
[15] T. Jacob, J. Wahr, W.T. Pfeffer, and S. Swenson. Recent contributions of glaciers and ice1045
caps to sea level rise. Nature, 482:514–518, 2012. doi: 10.1038/nature10847.1046
[16] A.S. Gardner, G. Moholdt, J. Cogley, B. Wouters, A.A. Arendt, J. Wahr, E. Berthier,1047
R. Hock, W.T. Pfeffer, G. Kaser, S.R.M. Ligtenberg, T. Bolch, M.J. Sharp, J.O. Hagen,1048
M.R. van den Broeke, and F. Paul. A Reconciled Estimate of Glacier Contributions to Sea1049
Level Rise: 2003 to 2009. Science, 340(6134):852–857, 2013. doi: 10.1126/science.1234532.1050
[17] Shuang Yi and Wenke Sun. Evaluation of glacier changes in high-mountain Asia based on1051
10year GRACE RL05 models. Journal of Geophysical Research: Solid Earth, 119(3):2504–1052
2517, 2014. doi: 10.1002/2013JB010860.1053
[18] A. Guntner. Improvement of global hydrological models using GRACE data. Surveys in1054
Geophysics, 29(4-5):375–397, 2008. doi: 10.1007/s10712-008-9038-y.1055
[19] B.R. Scanlon, L. Longuevergne, and D. Long. Ground referencing GRACE satellite es-1056
timates of groundwater storage changes in the California Central Valley, USA. Water1057
Resources Research, 48(4):W04520, 2012. doi: 10.1029/2011WR011312.1058
[20] J. Roberts and T.D. Roberts. Use of the butterworth low-pass filter for oceano-1059
graphic data. Journal of Geophysical Research: Oceans, 83(C11):5510–5514, 1978.1060
doi:10.1029/JC083iC11p05510.1061
[21] C. Sakumura, S. Bettadpur, and S. Bruinsma. Ensemble prediction and intercompari-1062
son analysis of GRACE time-variable gravity field models. Geophysical Research Letters,1063
41(5):1389–1397, 2014. doi: 10.1002/2013GL058632.1064
[22] H. Steffen, S. Petrovic, J. Muller, R. Schmidt, J. Wunsch, F. Barthelmes, and J. Kusche.1065
Significance of secular trends of mass variations determined from GRACE solutions. Jour-1066
nal of Geodynamics, 48(3-5):157–165, 2009. doi: 10.1016/j.jog.2009.09.029.1067
[23] V.R. Barletta R., L. Sandberg Sørensen, and R. Forsberg. Scatter of mass changes esti-1068
mates at basin scale for Greenland and Antarctica. The Cryosphere, 7(5):1411–1432, 2013.1069
doi: 10.5194/tc-7-1411-2013.1070
[24] D.P. Chambers and J.A. Bonin. Evaluation of Release-05 GRACE time-variable gravity1071
coefficients over the ocean. Ocean Science, 8(5):859–868, 2012. doi: 10.5194/os-8-859-2012.1072
[25] H. Save, S. Bettadpur, and B.D. Tapley. Reducing errors in the GRACE gravity solutions1073
using regularization. Journal of Geodesy, 86(9):695–711, 2012. doi: 10.1007/s00190-012-1074
0548-5.1075
[26] C. Jekeli. Alternative methods to smooth the Earth’s gravity field. Report No. 327. Ohio1076
State University Department of Geodetic Science and Surveying, 1981.1077
[27] S. Swenson and J. Wahr. Post-processing removal of correlated errors in GRACE data.1078
Geophysical Research Letters, 33(8):L08402, 2006. doi: 10.1029/2005GL025285.1079
[28] X.J. Duan, J.Y. Guo, C.K. Shum, and W. van der Wal. On the postprocessing removal1080
of correlated errors in GRACE temporal gravity field solutions. Journal of Geodesy,1081
83(11):1095–1106, 2009. doi: 10.1007/s00190-009-0327-0.1082
49
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 50: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/50.jpg)
[29] J. Kusche. Approximate decorrelation and nonisotropic smoothing of time-variable1083
GRACE-type gravity field models. Journal of Geodesy, 81(11):733–749, 2007. doi:1084
10.1007/s00190-007-0143-3.1085
[30] J. Kusche, R. Schmidt, S. Petrovic, and R. Rietbroek. Decorrelated GRACE time-variable1086
gravity solutions by GFZ, and their validation using a hydrological model. Journal of1087
Geodesy, 83(10):903–913, 2009. doi: 10.1007/s00190-009-0308-3.1088
[31] S. Werth, A. Guntner, R. Schmidt, and J. Kusche. Evaluation of grace filter tools from a1089
hydrological perspective. Geophysical Journal International, 179(3):1499–1515, 2009. doi:1090
10.1111/j.1365-246X.2009.04355.x.1091
[32] J.-M. Lemoine, S. Bruinsma, S. Loyer, R. Biancale, J.-C. Marty, F. Perosanz, and1092
G. Balmino. Temporal gravity field models inferred from GRACE data. Advances in1093
Space Research, 39(10):1620–1629, 2007. doi: 10.1016/j.asr.2007.03.062.1094
[33] S. Bruinsma, J.-M. Lemoine, R. Biancale, and N. Vales. CNES/GRGS 10-day gravity field1095
models (release 2) and their evaluation. Advances in Space Research, 45(4):587–601, 2010.1096
doi: 10.1016/j.asr.2009.10.012.1097
[34] J. L. Chen, Matt Rodell, C. R. Wilson, and J. S. Famiglietti. Low degree spherical har-1098
monic influences on Gravity Recovery and Climate Experiment (GRACE) water storage es-1099
timates. Geophysical Research Letters, 32(14):L14405, 2005. doi: 10.1029/2005GL022964.1100
[35] R. Rietbroek, M. Fritsche, S.-E. Brunnabend, I. Daras, J. Kusche, J. Schroter, F. Flecht-1101
ner, and R. Dietrich. Global surface mass from a new combination of GRACE, modelled1102
OBP and reprocessed GPS data. Journal of Geodynamics, 5960:64 – 71, 2012. doi:1103
10.1016/j.jog.2011.02.003.1104
[36] M. Cheng and B.D. Tapley. Variations in the earth’s oblateness during the past 281105
years. Journal of Geophysical Research: Solid Earth, 109(B9):B09402, 2004. doi:1106
10.1029/2004JB003028.1107
[37] G. Zhang, T. Yao, H. Xie, S. Kang, and Y. Lei. Increased mass over the Tibetan Plateau:1108
From lakes or glaciers? Geophysical Research Letters, 40(10):2125–2130, 2013. doi:1109
10.1002/grl.50462.1110
[38] L. Longuevergne, C. R. Wilson, B. R. Scanlon, and J. F. Cretaux. GRACE water storage1111
estimates for the Middle East and other regions with significant reservoir and lake storage.1112
Hydrology and Earth System Sciences, 17(12):4817–4830, 2013. soi: 10.5194/hess-17-4817-1113
2013.1114
[39] B. Lehner and P. Doll. Development and validation of a global database of lakes, reservoirs1115
and wetlands. Journal of Hydrology, 296(1):1–22, 2004. doi: 10.1016/j.jhydrol.2004.03.028.1116
[40] J.-F. Cretaux, W. Jelinski, S. Calmant, A. Kouraev, V. Vuglinski, M. Berge-Nguyen, M.-1117
C. Gennero, F. Nino, R. Abarca Del Rio, A. Cazenave, and P. Maisongrande. SOLS:1118
A lake database to monitor in the Near Real Time water level and storage variations1119
from remote sensing data. Advances in Space Research, 47(9):1497–1507, 2011. doi:1120
10.1016/j.asr.2011.01.004.1121
[41] C. Song, B. Huang, and L. Ke. Modeling and analysis of lake water storage changes on1122
the Tibetan Plateau using multi-mission satellite data. Remote Sensing of Environment,1123
135:25 – 35, 2013. doi: 10.1016/j.rse.2013.03.013.1124
50
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 51: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/51.jpg)
[42] L. Gao, J. Liao, and G. Shen. Monitoring lake-level changes in the Qinghai-Tibetan1125
Plateau using radar altimeter data (20022012). Journal of Applied Remote Sensing,1126
7(1):073470, 2013. doi: 10.1117/1.JRS.7.073470.1127
[43] V.B. Aizen, E.M. Aizen, and J.M. Melack. Climate, snow cover, glaciers, and runoff1128
in the Tien Shan, Central Asia. Journal of the American Water Resources Association,1129
31(6):1113–1129, 1995. doi: 10.1111/j.1752-1688.1995.tb03426.x.1130
[44] A. Paulson, S. Zhong, and J. Wahr. Limitations on the inversion for mantle viscosity from1131
postglacial rebound. Geophysical Journal International, 168(3):1195–1209, 2007. doi:1132
10.1111/j.1365-246X.2006.03222.x.1133
[45] Y. Yokoyama, K. Lambeck, P. De Deckker, P. Johnston, and L.K. Fifield. Timing of the1134
Last Glacial Maximum from observed sea-level minima. Nature, 406:713–716, 2000. doi:1135
10.1038/35021035.1136
[46] R.S Bradley and P.D. Jonest. ”Little Ice Age” summer temperature variations: their1137
nature and relevance to recent global warming trends. The Holocene, 3(4):367–376, 1993.1138
doi: 10.1177/095968369300300409.1139
[47] A. Geruo, J. Wahr, and S. Zhong. Computations of the viscoelastic response of a 3-D1140
compressible Earth to surface loading: an application to Glacial Isostatic Adjustment in1141
Antarctica and Canada. Geophysical Journal International, 192(2):557–572, 2012. doi:1142
10.1093/gji/ggs030.1143
[48] K. Lambeck, A. Purcell, J. Zhao, and N.-O. Svensson. The Scandinavian Ice Sheet:1144
from MIS 4 to the end of the Last Glacial Maximum. Boreas, 39(2):410–435, 2010. doi:1145
10.1111/j.1502-3885.2010.00140.x.1146
[49] M. Kuhle. The pleistocene glaciation of Tibet and the onset of ice ages An autocycle1147
hypothesis. GeoJournal, 17(4):581–595, 1988. doi: 10.1007/BF00209444.1148
[50] Y. Shi, B. Zheng, and S. Li. Last glaciation and maximum glaciation in the Qinghai-Xizang1149
(Tibet) Plateau: A controversy to M. Kuhle’s ice sheet hypothesis. Chinese Geographical1150
Science, 2(4):293–311, 1992. doi: 10.1007/BF02664561.1151
[51] J. Heyman, A.P. Stroeven, H. Alexanderson, C. Hattestrand, J. Harbor, Y. Li, M.W.1152
Caffee, L. Zhou, D. Veres, F. Liu, and M. Machiedo. Palaeoglaciation of Bayan Har1153
Shan, northeastern Tibetan Plateau: glacial geology indicates maximum extents limited1154
to ice cap and ice field scales. Journal of Quaternary Science, 24(7):710–727, 2009. doi:1155
10.1002/jqs.1305.1156
[52] Z. Li, K. Li, and L. Wang. Study on recent glacier changes and their impact on water1157
resources in Xinjiang, Northwestern China. Quaternary Sciences, 30(1):96–106, 2010. doi:1158
10.3969/j.issn.1001-7410.2010.01.09.1159
[53] S.D. Willett. Orogeny and orography: The effects of erosion on the structure of mountain1160
belts. Journal of Geophysical Research: Solid Earth, 104(B12):28957–28981, 1999. doi:1161
10.1029/1999JB900248.1162
[54] A.V. Zubovich, X.-Q. Wang, Y.G. Scherba, G.G. Schelochkov, R. Reilinger, C. Reigber,1163
O.I. Mosienko, P. Molnar, W. Michajljow, V.I. Makarov, J. Li, S.I. Kuzikov, T.A. Herring,1164
M.W. Hamburger, B.H. Hager, Y.-M. Dang, V.D. Bragin, and R.T. Beisenbaev. GPS1165
velocity field for the Tien Shan and surrounding regions. Tectonics, 29(6):TC6014, 2010.1166
doi: 10.1029/2010TC002772.1167
51
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 52: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/52.jpg)
[55] W. Sun, Q. Wang, H. Li, Y. Wang, S. Okubo, D. Shao, D. Liu, and G. Fu. Gravity1168
and GPS measurements reveal mass loss beneath the Tibetan Plateau: Geodetic evidence1169
of increasing crustal thickness. Geophysical Research Letters, 36(2):L02303, 2009. doi:1170
10.1029/2008GL036512.1171
[56] E.V. Burov, M.G. Kogan, Hlne Lyon-Caen, and Peter Molnar. Gravity anomalies, the deep1172
structure, and dynamic processes beneath the Tien Shan. Earth and Planetary Science1173
Letters, 96(34):367 – 383, 1990. doi: 10.1016/0012-821X(90)90013-N.1174
[57] D.D. Rowlands, S.B. Luthcke, S.M. Klosko, F.G.R. Lemoine, D.S. Chinn, J.J. McCarthy,1175
C.M. Cox, and O.B. Anderson. Resolving mass flux at high spatial and temporal resolution1176
using GRACE intersatellite measurements. Geophysical Research Letters, 32:L04310, 2005.1177
doi: 10.1029/2004GL021908.1178
[58] A. Eicker, J. Schall, and J. Kusche. Regional gravity modelling from spaceborne data:1179
case studies with GOCE. Geophysical Journal International, 196(3):1431–1440, 2014. doi:1180
10.1093/gji/ggt485.1181
[59] B. Wouters, Chambers, and E.J.O. Schrama. GRACE observes small-scale mass loss in1182
Greenland. Geophysical Research Letters, 35:L20501, 2008. doi: 10.1029/2008GL034816.1183
[60] J.L. Chen, C.R. Wilson, and B.D. Tapley. Interannual variability of Greenland ice losses1184
from satellite gravimetry. Journal of Geophysical Research: Solid Earth, 116(B7):B07406,1185
2011. doi: 10.1029/2010JB007789.1186
[61] Oliver Baur and Nico Sneeuw. Assessing Greenland ice mass loss by means of point-1187
mass modeling: a viable methodology. Journal of Geodesy, 85(9):607–615, 2011. doi:1188
10.1007/s00190-011-0463-1.1189
[62] E.J.O. Schrama, B. Wouters, and R. Rietbroek. A mascon approach to assess ice sheet and1190
glacier mass balances and their uncertainties from GRACE data. Journal of Geophysical1191
Research: Solid Earth, 119(7):6048–6066, 2014. doi: 10.1002/2013jb010923.1192
[63] M. Horwath and R. Dietrich. Signal and error in mass change inferences from GRACE:1193
the case of Antarctica. Geophysical Journal International, 177(3):849–864, 2009. doi:1194
10.1111/j.1365-246X.2009.04139.x.1195
[64] W. Colgan, S. Luthcke, W. Abdalati, and M. Citterio. Constraining GRACE-derived1196
cryosphere-attributed signal to irregularly shaped ice-covered areas. The Cryosphere,1197
7(6):1901–1914, 2013. doi: 10.5194/tc-7-1901-2013.1198
[65] W. Feng, M. Zhong, J.-M. Lemoine, R. Biancale, H.-T. Hsu, and J. Xia. Evaluation1199
of groundwater depletion in North China using the Gravity Recovery and Climate Ex-1200
periment (GRACE) data and ground-based measurements. Water Resources Research,1201
49:2110–2118, 2013. doi: 10.1002/wrcr.20192.1202
[66] A. Tarantola and B. Valette. Inverse problems = quest for information. Journal of Geo-1203
physics, 50:159–170, 1982.1204
[67] W.T. Pfeffer and the Randolph Consortium. The Randolph Glacier Inventory: a glob-1205
ally complete inventory of glaciers. Journal of Glaciology, 60(221):537–552, 2014. doi:1206
10.3189/2014JoG13J176.1207
[68] J. Wahr, S. Swenson, V. Zlotnicki, and I. Velicogna. Time-variable gravity from1208
GRACE: First results. Geophysical Research Letters, 31(11):L11501, 2004. doi:1209
10.1029/2004GL019779.1210
52
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 53: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/53.jpg)
[69] J.L. Chen, C.R. Wilson, and K.W. Seo. S2 tide aliasing in GRACE time-variable gravity1211
solutions. Journal of Geodesy, 83:679–687, 2009. doi: 10.1007/s00190-008-0282-1.1212
[70] S.V. Bettadpur. GRACE UTCSR Level-2 processing standards document (for product1213
Release 0005). Technical report, Center for Space Research - University of Texas, U.S.A.,1214
2012.1215
[71] C. Dahle, F. Flechtner, C. Gruber, D. Konig, R. Konig, G. Michalak, and K.-H. Neumayer.1216
GFZ GRACE Level-2 processing standards document for level-2 product Release 0005.1217
Scientific technical report STR12/02 - Data, GFZ German Research Centre for Geosciences1218
- Potsdam, Germany, 2012. doi: 10.2312/GFZ.b103-12020.1219
[72] M.M. Watkins and D.-N. Yuan. GRACE JPL Level-2 processing standards document (for1220
product Release 05). Technical report, Jet Propulsion Laboratory - Pasadena, U.S.A.,1221
2012.1222
[73] M. Rodell, P.R. Houser, U. Jambor, J. Gottschalck, K. Mitchell, C.-J. Meng, K. Arsenault,1223
B. Cosgrove, J. Radakovich, M. Bosilovich, J.K. Entin, J.P. Walker, D. Lohmann, and1224
D. Toll. The global land data assimilation system. Bulletin of the American Meteorological1225
Society, 85(3):381–394, 2004. doi: 10.1175/BAMS-85-3-381.1226
[74] K.W. Oleson, D.M. Lawrence, G.B. Bonan, M.G. Flanner, E. Kluzek, J. Peter, S. Levis,1227
S.C. Swenson, E. Thornton, J. Feddema, C.L. Heald, J.-F. Lamarque, G.-Y. Niu, T. Qian,1228
S. Running, K. Sakaguchi, L. Yang, X. Zeng, and X. Zeng. Technical description of version1229
4.0 of the Community Land Model (CLM). NCAR Technical Note, NCAR/TN-478+STR,1230
2010. ISSN: 2153-2400.1231
[75] K.W. Oleson, D.M. Lawrence, G.B. Bonan, B. Drewniak, M. Huang, C.D. Koven, S. Levis,1232
F. Li, W.J. Riley, Z.M. Subin, S.C. Swenson, and P.E. Thornton. Technical description of1233
version 4.5 of the Community Land Model (CLM). NCAR Technical Note, NCAR/TN-1234
503+STR, 2013. ISSN: 2153-2400.1235
[76] J. Sheffield, G. Goteti, and E.F. Wood. Development of a 50-year high-resolution global1236
dataset of meteorological forcings for land surface modeling. Journal of Climate, 19:3088–1237
3111, 2006. doi: 10.1175/JCLI3790.1.1238
[77] P. Doll, F. Kaspar, and L. Lehner. A global hydrological model for deriving water avail-1239
ability indicators: model tuning and validation. Journal of Hydrology, 270(1-2):105–134,1240
2003. doi: 10.1016/S0022-1694(02)00283-4.1241
[78] H. Muller Schmied, S. Eisner, D. Franz, M. Wattenbach, F. T. Portmann, M. Florke,1242
and P. Doll. Sensitivity of simulated global-scale freshwater fluxes and storages to input1243
data, hydrological model structure, human water use and calibration. Hydrology and Earth1244
System Sciences Discussions, 11(2):1583–1649, 2014. doi: 10.5194/hessd-11-1583-2014.1245
[79] A.I.J.M. van Dijk, J.L Pena-Arancibia, E.F. Wood, J. Sheffield, and H.E. Beck. Global1246
analysis of seasonal streamflow predictability using an ensemble prediction system and ob-1247
servations from 6192 small catchments worldwide. Water Resources Research, 49(5):2729–1248
2746, 2013. doi: 10.1002/wrcr.20251.1249
[80] C.E. Webb, H.J. Zwally, and W. Abdalati. The Ice, Cloud and land Elevation Satellite1250
(ICESat): Summary mission timeline and performance relative to pre-launch mission suc-1251
cess criteria. Technical Report NASA/TM-2013-217512, National Aeronautics and Space1252
Administration, Goddard Space Flight Center Greenbelt, Maryland, 2013.1253
53
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 54: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/54.jpg)
[81] H.J. Zwally, B. Schutz, W. Abdalati, J. Abshire, C. Bentley, A. Brenner, J. Bufton,1254
J. Dezio, D. Hancock, D. Harding, T. Herring, B. Minster, K. Quinn, S. Palm, J. Spinhirne,1255
and R. Thomas. ICESat’s laser measurements of polar ice, atmosphere, ocean, and land.1256
Journal of Geodynamics, 34(3-4):405–444, 2002. doi: 10.1016/S0264-3707(02)00042-X.1257
[82] J.B. Abshire, X. Sun, H. Riris, , J.M. Sirota, J.F. McGarry, S. Palm, D. Yi, and1258
P. Liiva. Geoscience Laser Altimeter System (GLAS) on the ICESat Mission: On-1259
orbit measurement performance. Geophysical Research Letters, 32:L21S02, 2005. doi:1260
10.1029/2005GL024028.1261
[83] X. Wang, X. Cheng, P. Gonga, H. Huang, Z. Li, and X. Li. Earth science applications1262
of ICESat/GLAS: a review. International Journal of Remote Sensing, 32(23):8837–8864,1263
2011. doi: 10.1080/01431161.2010.547533.1264
[84] A. Kaab, E. Berthier, C. Nuth, J. Gardelle, and Y. Arnaud. Contrasting patterns of early1265
twenty-first-century glacier mass change in the Himalayas. Nature, 488:495–498, 2012. doi:1266
10.1038/nature11324.1267
[85] N. Neckel, J. Kropacek, T. Bolch, and V. Hochschild. Glacier mass changes on the Tibetan1268
Plateau 2003-2009 derived from ICESat laser altimetry measurements. Environmental1269
Research Letters, 9(1):014009, 2014. doi: 10.1088/1748-9326/9/1/014009.1270
[86] H. Zwally, R. Schutz, C. Bentley, J. Bufton, T. Herring, J. Minster, J. Spinhirne, and1271
R. Thomas. GLAS/ICESat L1B Global Elevation Data. Version 33. National Snow and1272
Ice Data Center, Boulder, Colorado, USA, 2011. Digital Media.1273
[87] H.D. Pritchard, R.J. Arthern1, D.G. Vaughan, and L.A. Edwards. Extensive dynamic1274
thinning on the margins of the Greenland and Antarctic ice sheets. Nature, 461:971–975,1275
2009. doi: 10.1038/nature08471.1276
[88] A.S. Gardner, G. Moholdt, B. Wouters, G.J. Wolken, D.O. Burgess, M.J. Sharp, J.G.1277
Cogley, C. Braun, and C. Labine. Sharply increased mass loss from glaciers and ice caps1278
in the Canadian Arctic Archipelago. Nature, 473:357–360, 2011. doi: 10.1038/nature10089.1279
[89] G. Moholdt, C. Nuth, J.O. Hagen, and J. Kohler. Recent elevation changes of Sval-1280
bard glaciers derived from ICESat laser altimetry. Remote Sensing of Environment,1281
114(11):2756–2767, 2010. doi: 10.1016/j.rse.2010.06.008.1282
[90] N.K. Pavlis, S.A. Holmes, S.C. Kenyon, and J.K. Factor. The development and evaluation1283
of the Earth Gravitational Model 2008 (EGM2008). Journal of Geophysical Research:1284
Solid Earth, 117:B04406, 2012. doi: 10.1029/2011JB008916.1285
[91] R. S. Nerem, F. J. Lerch, J. A. Marshall, E. C. Pavlis, B. H. Putney, B. D. Tapley, R. J.1286
Eanes, J. C. Ries, B. E. Schutz, C. K. Shum, M. M. Watkins, S. M. Klosko, J. C. Chan,1287
S. B. Luthcke, G. B. Patel, N. K. Pavlis, R. G. Williamson, R. H. Rapp, R. Biancale, and1288
F. Nouel. Gravity model development for TOPEX/POSEIDON: Joint Gravity Models1289
1 and 2. Journal of Geophysical Research: Oceans, 99(C12):24421–24447, 1994. doi:1290
10.1029/94JC01376.1291
[92] H.A. Fricker, A. Borsa, B. Minster, C. Carabajal, K. Quinn, and B. Bills. Assessment1292
of ICESat performance at the salar de Uyuni, Bolivia. Geophysical Research Letters,1293
32:L21S06, 2005. doi: 10.1029/2005GL023423.1294
54
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 55: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/55.jpg)
[93] J. Jarvis, H. Reuter, A. Nelson, and E. Guevara. Hole-filled SRTM for the globe. CGIAR1295
Consortium for Spatial Information, Montpellier, France, 2008. CGIAR-CSI SRTM 90 m1296
Database, Version 4, http://srtm.csi.cgiar.org/.1297
[94] E. Rodriguez, C.S. Morris, and J.E. Belz. A global assessment of the SRTM performance.1298
Photogrammetric Engineering and Remote Sensing, 72(3):249–260, 2006.1299
[95] J. Gardelle, E. Berthier, and Y. Arnaud. Slight mass gain of Karakoram glaciers in the1300
early twenty-first century. Nature Geoscience, 5:322–325, 2012. doi: 10.1038/ngeo1450.1301
[96] P.J. Rousseeuw and A.M. Leroy. Robust Regression and Outlier Detection. Wiley series1302
in probability and statistics. John Wiley & Sons, 2005.1303
[97] F. Mosteller and J.W. Tukey. Data analysis and regression: A second course in statistics.1304
Addison-Wesley, Reading, MA, USA, 1977.1305
[98] M. Huss. Density assumptions for converting geodetic glacier volume change to mass1306
change. The Cryosphere, 7(3):877–887, 2013. doi: 10.5194/tc-7-877-2013.1307
[99] G. Moholdt, B. Wouters, and A.S. Gardner. Recent mass changes of glaciers in1308
the Russian High Arctic. Geophysical Research Letters, 39(10):L10502, 2012. doi:1309
10.1029/2012GL051466.1310
[100] M.A. Hofton, S.B. Luthcke, and J.B. Blair. Estimation of ICESat intercampaign elevation1311
biases from comparison of lidar data in East Antarctica. Geophysical Research Letters,1312
40(21):5698–5703, 2013. doi: 10.1002/2013GL057652.1313
[101] D.W. Burbank, J. Leland, E. Fielding, R.S. Anderson, N. Brozovic, M.R. Reid, and1314
C. Duncan. Bedrock incision, rock uplift and threshold hillslopes in the northwestern1315
Himalayas. Nature, 379(6565):505–510, 1996. doi:10.1038/379505a0.1316
[102] A. Arendt, , and 50 others. Randolph Glacier Inventory [3.2]: A Dataset of Global Glacier1317
Outlines. Global Land Ice Measurements from Space (GLIMS), Boulder Colorado, USA,1318
2013. Digital Media.1319
[103] J.S. Kargel, G.J. Leonard, M.P. Bishop, A. Kaab, and B.H. Raup, editors. Global Land1320
Ice Measurements from Space. Springer Praxis Books. Springer, 2014.1321
[104] Y. Shi, C. Liu, and E. Kang. The Glacier Inventory of China. Annals of Glaciology,1322
50(53):1–4, 2010. doi:10.3189/172756410790595831.1323
[105] T. Bolch. Climate change and glacier retreat in northern Tien Shan (Kaza-1324
khstan/Kyrgyzstan) using remote sensing data. Global and Planetary Change, 56(1):1–12,1325
2007. doi: 10.1016/j.gloplacha.2006.07.009.1326
[106] S. Kutuzov and M. Shahgedanova. Glacier retreat and climatic variability in the east-1327
ern TerskeyAlatoo, inner Tien Shan between the middle of the 19th century and be-1328
ginning of the 21st century. Global and Planetary Change, 69(1):59–70, 2009. doi:1329
10.1016/j.gloplacha.2009.07.001.1330
[107] D. Kriegel, C. Mayer, W. Hagg, S. Vorogushyn, D. Duethmann, A. Gafurov, and1331
D. Farinotti. Changes in glacierisation, climate and runoff in the second half of the 20th1332
century in the Naryn basin, Central Asia. Global and Planetary Change, 110(part A):51–1333
56, 2013. doi: 10.1016/j.gloplacha.2013.05.014.1334
55
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 56: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/56.jpg)
[108] A. Osmonov, T. Bolch, C. Xi, A. Kurban, and W. Guo. Glacier characteristics and changes1335
in the Sary-Jaz River Basin (Central Tien Shan, Kyrgyzstan) - 1990-2010. Remote Sensing1336
Letters, 4(8):725–734, 2013. doi: 10.1080/2150704X.2013.789146.1337
[109] B.H. Raup, H.H. Kieffer, T.M. Hare, and J.S. Kargel. Generation of data acquisition1338
requests for the ASTER satellite instrument for monitoring a globally distributed target:1339
glaciers. IEEE Transactions on Geoscience and Remote Sensing, 38(2):1105–1112, 2000.1340
doi: 10.1109/36.841989.1341
[110] D.M. Danko. The digital chart of the world project. Photogrammetric Engineering and1342
Remote Sensing, 58(8):1125–1128, 1992.1343
[111] W. Haeberli, M. Hoelzle, S. Suter, and World Glacier Monitoring Service. Into the second1344
century of worldwide glacier monitoring - Prospects and strategies: A contribution to the1345
International Hydrological Programme (IHP) and the Global Environment Monitoring Sys-1346
tem (GEMS). Studies and Reports in Hydrology. Renouf Publishing Company Limited,1347
1998.1348
[112] J.G. Cogley. Present and future states of Himalaya and Karakoram glaciers. Annals of1349
Glaciology, 52(59):69–73, 2011. doi: 10.3189/172756411799096277.1350
[113] S.M. Uppala, , and 45 others. The ERA-40 re-analysis. Quarterly Journal of the Royal1351
Meteorological Society, 131(612):2961–3012, 2006. doi: 10.1256/qj.04.176.1352
[114] D.P. Dee, , and 35 others. The ERA-Interim reanalysis: configuration and performance1353
of the data assimilation system. Quarterly Journal of the Royal Meteorological Society,1354
137(656):553–597, 2011. doi: 10.1002/qj.828.1355
[115] E. Kalnay, M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, M. Iredell,1356
S. Saha, G. White, J. Woollen, Y. Zhu, A. Leetmaa, and R. Reynolds. The NCEP/NCAR1357
40-year reanalysis project. Bulletin of the American Meteorological Society, 77:437–472,1358
1996. doi: 10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.1359
[116] M.J. Menne, I. Durre, R.S. Vose, B.E. Gleason, and T.G. Houston. An Overview of1360
the Global Historical Climatology Network-Daily Database. Journal of Atmospheric and1361
Oceanic Technology, 29(7):897–910, 2012. doi: 10.1175/jtech-d-11-00103.1.1362
[117] M. Huss, A. Bauder, M. Funk, and R. Hock. Determination of the seasonal mass balance1363
of four Alpine glaciers since 1865. Journal of Geophysical Research, 113:F01015, 2008. doi:1364
10.1029/2007JF000803.1365
[118] D. Farinotti, S. Usselmann, M. Huss, A. Bauder, and M. Funk. Runoff evolution in1366
the Swiss Alps: Projections for selected high-alpine catchments based on ENSEMBLES1367
scenarios. Hydrological Processes, 26(13):1909–1924, 2012. doi: 10.1002/hyp.8276.1368
[119] E. L. Peck and M. J. Brown. An approach to the development of isohyetal maps1369
for mountainous areas. Journal of Geophysical Research, 67:681–694, 1962. doi:1370
10.1029/JZ067i002p00681.1371
[120] D. G. Tarboton, T. G. Chowdhury, and T. H. Jackson. A spatially distributed energy1372
balance snowmelt model. In K. A. Tonnessen, M. W. Williams, and M. Tranter, editors,1373
Biogeochemistry of seasonally snow-covered catchments, Proceedings of a Boulder Sympo-1374
sium, pages 141–155. IAHS Publ. No. 228, 1995.1375
56
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 57: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/57.jpg)
[121] M. Huss, A. Bauder, and M. Funk. Homogenization of long-term mass balance time series.1376
Annals of Glaciology, 50(50):198–206, 2009. doi: 10.3189/172756409787769627.1377
[122] R. Hock. A distributed temperature-index ice- and snowmelt model including potential1378
direct solar radiation. Journal of Glaciology, 45(149):101–111, 1999.1379
[123] M. Huss, R. Hock, A. Bauder, and M. Funk. 100-year glacier mass changes in the1380
Swiss Alps linked to the Atlantic Multidecadal Oscillation. Geophysical Research Letters,1381
37:L10501, 2010. doi: 10.1029/2010GL042616.1382
[124] M. Huss, D. Farinotti, A. Bauder, and M. Funk. Modelling runoff from highly glacierized1383
alpine catchment basins in a changing climate. Hydrological Processes, 22(19):3888–3902,1384
2008. doi: 10.1002/hyp.7055.1385
[125] M. Huss, S. Usselmann, D. Farinotti, and A. Bauder. Glacier mass balance in the south-1386
eastern Swiss Alps since 1900 and perspectives for the future. Erdkunde, 64(2):119–140,1387
2010. doi: 10.3112/erdkunde.2010.02.02.1388
[126] T. Molg and G. Kaser. A new approach to resolving climate-cryosphere relations: Down-1389
scaling climate dynamics to glacier-scale mass and energy balance without statistical scale1390
linking. Journal of Geophysical Research: Atmospheres, 116(D16):D16101, 2011. doi:1391
10.1029/2011JD015669.1392
[127] J. Oerlemans. Glaciers and climate change. A.A. Balkema Publishers, Lisse, Netherlands,1393
2001. 148 pp.1394
[128] A.S. Gardner and M.J. Sharp. A review of snow and ice albedo and the development1395
of a new physically based broadband albedo parameterization. Journal of Geophysical1396
Research, 115:F01009, 2010. doi: 10.1029/2009JF001444.1397
[129] T. Pieczonka and T. Bolch. Region-wide glacier mass budgets and area changes for the1398
Central Tien Shan between 1975 and 1999 using Hexagon KH-9 imagery. Global and1399
Planetary Change, 128:1–13, 2015. doi: 10.1016/j.gloplacha.2014.11.014.1400
[130] T. Pieczonka, T. Bolch, W. Junfeng, and L. Shiyin. Heterogeneous mass loss of glaciers1401
in the Aksu-Tarim Catchment (Central Tien Shan) revealed by 1976 KH-9 Hexagon and1402
2009 SPOT-5 stereo imagery. Remote Sensing of Environment, 130:233–244, 2013. doi:1403
10.1016/j.rse.2012.11.020.1404
[131] T. Bolch, T. Pieczonka, and D.I. Benn. Multi-decadal mass loss of glaciers in the Everest1405
area (Nepal Himalaya) derived from stereo imagery. The Cryosphere, 5(2):349–358, 2011.1406
doi: 10.5194/tc-5-349-2011.1407
[132] T. Nuimura, K. Fujita, S. Yamaguchi, and R.R. Sharma. Elevation changes of glaciers1408
revealed by multitemporal digital elevation models calibrated by GPS survey in the1409
Khumbu region, Nepal Himalaya, 1992-2008. Journal of Glaciology, 58(210):648–656,1410
2012. doi:10.3189/2012JoG11J061.1411
[133] M. Huss. Extrapolating glacier mass balance to the mountain-range scale: the European1412
Alps 1900-2100. The Cryosphere, 6:713–727, 2012. doi: 10.5194/tc-6-713-2012.1413
[134] J.G. Cogley, R. Hock, L.A. Rasmussen, A.A. Arendt, A. Bauder, R.J. Braithwaite, P. Jans-1414
son, G. Kaser, M. Moller, L. Nicholson, and M. Zemp. Glossary of glacier mass balance1415
and related terms. IHP-VII Technical Documents in Hydrology No. 86, IACS Contribution1416
No. 2, UNESCO-IHP, Paris, 2011.1417
57
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 58: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/58.jpg)
[135] D.H. Elsberg, W.D. Harrison, K.A. Echelmeyer, and R.M. Krimmel. Quantifying the1418
effects of climate and surface change on glacier mass balance. Journal of Glaciology,1419
47(159):649–658, 2001. doi: 10.3189/172756501781831783.1420
[136] W.D. Harrison, D.H. Elsberg, L.H. Cox, and R.S. March. Correspondence. Different mass1421
balances for climatic and hydrologic applications. Journal of Glaciology, 51(172):176, 2005.1422
doi: 10.3189/172756505781829601.1423
[137] W.D. Harrison, L.H. Cox, R. Hock, R.S. March, and E.C. Pettit. Implications for the dy-1424
namic health of a glacier from comparison of conventional and reference-surface balances.1425
Annals of Glaciology, 50(50):25–30, 2009. doi: 10.3189/172756409787769654.1426
[138] M. Huss, R. Hock, A. Bauder, and M. Funk. Conventional versus reference-surface mass1427
balance. Journal of Glaciology, 58(208):278–286, 2012. doi: 10.3189/2012JoG11J216.1428
[139] V.B. Aizen, V.A. Kuzmichenok, A.B. Surazakov, and E.M. Aizen. Glacier changes1429
in the central and northern Tien Shan during the last 140 years based on sur-1430
face and remote-sensing data. Annals of Glaciology, 43(1):202–213, 2006. doi:1431
10.3189/172756406781812465.1432
[140] G. Østrem and M. Brugman. Glacier mass-balance measurements - A manual for field1433
and office work. Technical report, National Hydrology Research Institute, 1991. Science1434
Report No. 4.1435
[141] C. Narama, A. Kaab, M. Duishonakunov, and K. Abdrakhmatov. Spatial variability1436
of recent glacier area changes in the Tien Shan Mountains, Central Asia, using Corona1437
(1970), Landsat (2000), and ALOS (2007) satellite data. Global and Planetary Change,1438
71(1):42–54, 2010. doi: 10.1016/j.gloplacha.2009.08.002.1439
[142] S. Wang, M. Zhang, Z. Li, F. Wang, H. Li, Y. Li, and X. Huang. Glacier area variation and1440
climate change in the Chinese Tianshan Mountains since 1960. Journal of Geographical1441
Sciences, 21(2):263–273, 2011. doi: 10.1007/s11442-011-0843-8.1442
[143] A.M. Breipohl. Probabilistic systems analysis: an introduction to probabilistic models,1443
decisions, and applications of random processes. Wiley, 1970.1444
[144] D.M. Allen. The relationship between variable selection and data augmenta-1445
tion and a method for prediction. Technometrics, 16(1):125–127, 1974. doi:1446
10.1080/00401706.1974.10489157.1447
[145] M. Huss and D. Farinotti. Distributed ice thickness and volume of all glaciers around the1448
globe. Journal of Geophysical Research, 117:F04010,, 2012. doi: 10.1029/2012JF002523.1449
[146] N.R. Draper, H. Smith, and E. Pownell. Applied regression analysis. Wiley-Interscience,1450
New York, 1998.1451
[147] E.W. Steyerberg, M.J.C. Eijkemansa, and J.D.F. Habbema. Stepwise selection in small1452
data sets: A simulation study of bias in logistic regression analysis. Journal of Clinical1453
Epidemiology, 52(10):935–942, 1999.1454
[148] M. Schatzoff, R. Tsao, and S. Fienberg. Efficient calculation of all possible regressions.1455
Technometrics, 10(4):769–779, 1968. doi: 10.1080/00401706.1968.10490629.1456
[149] L. Wilkinson. Revising the Pareto chart. The American Statistician, 60(4):332–334, 2006.1457
doi: 10.1198/000313006X152243.1458
58
© 2015 Macmillan Publishers Limited. All rights reserved
![Page 59: SUPPLEMENTARY INFORMATION - Nature...112 due to the orbital geometry and short separation between the GRACE satellites [34]. Here, the 113 impact of considering/ignoring the substitution](https://reader034.vdocuments.site/reader034/viewer/2022042114/5e906b7c3e2410684f0f3afd/html5/thumbnails/59.jpg)
[150] S.N. Ushnurtsev. Mass balance fluctuations of the Sary-Tor glacier in inner Tien Shan1459
and its reconstruction for the period 1930-1988. Glaciological Studies, 71:70–79, 1991. (in1460
Russian).1461
[151] World Glacier Monitoring Service (WGMS). Glacier mass balance bulletins -1462
Bulletins No. 1-12 (1988-1989, 2010-2011). ICSU (WDS) / IUGG (IACS) /1463
UNEP / UNESCO / WMO, Zurich, Switzerland, 1991-2012. Available online at1464
http://www.geo.uzh.ch/microsite/wgms/gmbb.html. Publication based on database ver-1465
sion: doi:10.5904/wgms-fog-2013-11.1466
[152] M.B. Dyurgerov and M.F. Meier. Glaciers and the changing earth system: A 2004 snap-1467
shot. Occasional Paper 58, Institute of Arctic and Alpine Research, University of Colorado,1468
2005. pp. 117.1469
[153] W. Hagg, C. Mayer, A. Lambrecht, D. Kriegel, and E. Azizov. Glacier changes in the Big1470
Naryn basin, Central Tian Shan. Global and Planetary Change, 110(Part A):40–50, 2013.1471
doi: 10.1016/j.gloplacha.2012.07.010.1472
[154] S. Liu, Y. Ding, D. Shangguan, Y. Zhang, J. Li, H. Han, J. Wang, and C. Xie.1473
Glacier retreat as a result of climate warming and increased precipitation in the1474
Tarim river basin, northwest China. Annals of Glaciology, 43(1):91–96, 2006. doi:1475
10.3189/172756406781812168.1476
[155] B. Li, A.-X. Zhu, Y. Zhang, T. Pei, C. Qin, and C Zhou. Glacier change over the past1477
four decades in the middle Chinese Tien Shan. Journal of Glaciology, 52(178):425–438,1478
2006. doi: 10.3189/17275650678182855.1479
[156] K.M. Li, Z.Q. Li, W.Y. Gao, and L. Wang. Recent glacial retreat and its effect on water1480
resources in eastern Xinjiang. Chinese Science Bulletin, 56(33):3596–3604, 2011. doi:1481
10.1007/s11434-011-4720-8.1482
[157] L. Wang, Z. Li, F. Wang, and R. Edwards. Glacier shrinkage in the Ebinur lake basin,1483
Tien Shan, China, during the past 40 years. Journal of Glaciology, 60(220):245–254, 2014.1484
doi: 10.3189/2014JoG13J023.1485
[158] C. Narama, Yuichi Shimamura, D Nakayama, and K. Abdrakhmatov. Recent changes of1486
glacier coverage in the western Terskey-Alatoo range, Kyrgyz Republic, using Corona and1487
Landsat. Annals of Glaciology, 43(1):223–229, 2006. doi: 10.3189/172756406781812195.1488
[159] P. Niederer, V. Bilenko, N. Ershova, H. Hurni, S. Yerokhin, and D. Maselli. Tracing1489
glacier wastage in the Northern Tien Shan (Kyrgyzstan/Central Asia) over the last 401490
years. Climatic Change, 86(1-2):227–234, 2008. doi: 10.1007/s10584-007-9288-6.1491
59© 2015 Macmillan Publishers Limited. All rights reserved