an improved logistic method for detecting spring vegetation.pdf
TRANSCRIPT
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 1/12
Agricultural and Forest Meteorology 200 (2015) 9–20
Contents lists available at ScienceDirect
Agricultural and Forest Meteorology
j ournal homepage : www.elsevier .com/ locate /agr formet
An improved logistic method for detecting spring vegetation
phenology in grasslands fromMODIS EVI time-series data
Ruyin Caoa, Jin Chenb,c,∗, Miaogen Shend, Yanhong Tang a
a National Institute for Environmental Studies, Onogawa 16-2, Tsukuba, Ibaraki, 305-8506, Japanb State Key Laboratory of Earth Surface Processesand Resource Ecology, Beijing NormalUniversity, Beijing, 100875, Chinac College of Global Changeand Earth SystemScience, Beijing NormalUniversity, Beijing, 100875, Chinad Key Laboratory of Alpine Ecology and Biodiversity, Institute of Tibetan Plateau Research,Chinese Academy of Sciences,Beijing, 100101, China
a r t i c l e i n f o
Article history:
Received 13March 2014
Received in revised form 9 September2014
Accepted 14 September2014
Keywords:
Climate change
Green up
Inner Mongolia
Logistic fitting
Precipitation
Start of the growing season
a b s t r a c t
Satellite-derived greenness vegetation indices provide a valuable data source for characterizing spring
vegetation phenology over regional or global scales. A logistic function has been widely used to fit time
series of vegetation indices to estimate green-update (GUD),which is currently beingused for generating
the global phenological product from the Enhanced Vegetation Index (EVI) time-series data provided
by the Moderate Resolution Imaging Spectroradiometer (MODIS). In this study, we address a violation
of the basic assumption of the logistic fitting method that arises from the fact that vegetation growth
under natural conditions is controlled by multiple environmental factors and often does not follow a
well-definedS-shaped logistic temporal profile.We developed the adaptive local iterative logistic fitting
method (ALILF) to analyze the “local range” (i.e., the range of data points where the values in the time
series begin to increase rapidly) in theMODIS EVI profile inwhichGUD is found. The newmethod adopts
an iterative procedure and an adaptive temporal window to properly simulate the trajectory of EVI time
series in the local range, and can determineGUDmore accurately. GUD estimated by ALILF almostmatch
the date of the onset of the greenness increase well while the traditional logistic fitting method shows
errors of even more than 1 month in the same cases. ALILF is a more general form of the logistic fitting
method that can estimate GUD both fromwell-defined S-shaped time series and from non-logistic ones.Besides, it is resistant to a range of noise levels added on the time-series data (Gaussian noise with a
meanvalueof zero and standard deviations ranging from 0% to 15% of the EVI value). These advantages
mean ALILF may be widely used for monitoring spring vegetation phenology from greenness vegetation
indices.
© 2014 Elsevier B.V. All rights reserved.
1. Introduction
Spring vegetation phenology refers to the onset of photosyn-
thetic activity, and is controlledbymultipleenvironmental factors.
Spring temperature has been widely accepted as the main factor
driving spring phenology in temperate forests (Piao et al., 2006,
Richardson et al., 2006), and spring precipitation is considered to
be a main driver for deserts and temperate grasslands (Cong et al.,
Abbreviations: ALILF, adaptive local iterative logistic fitting method; AVHRR,
advanced very high resolution radiometer; DOY, day of year; EVI, enhanced
vegetation index; GUD, green-update;MODIS, moderate resolution imaging spec-
troradiometer.∗ Corresponding author at: State Key Laboratory of Earth Surface Processes and
Resource Ecology, Beijing Normal University, Beijing, 100875, China.
Tel.: +86 13522889711.
E-mail addresses: [email protected], [email protected] (J. Chen).
2012;Shenet al., 2011). Other less obvious factors such asphotope-
riod (Partanen et al., 1998) also affect spring phenology. Recent
climate change, particularly spring warming, has greatly altered
spring vegetationphenology (e.g., Menzel et al., 2006; Jeong et al.,
2012). The changes in springphenologyareecologically important
because they strongly affect carbon cycling and energy balance
in terrestrial ecosystems (Chapin et al., 2008; Jeong et al., 2009;
Richardson et al., 2009). For instance,an earlier onset in springwas
found to be one of the main factors to increase the carbon sink
for northern hemisphere terrestrial ecosystems (Piao et al., 2008).
It is thus very significant to monitor spring vegetation phenology
to gain insights into linkage between phenology andclimate at the
largescale,whichis ahottopic inglobalchange research.Currently,
monitoring springphenologywith wall-to-wall spatial coverage is
only available based on satellite remotely sensed data.
Two main types of methods have been developed to determine
thetiming of springphenology from time seriesof satellite-derived
http://dx.doi.org/10.1016/j.agrformet.2014.09.009
0168-1923/©2014 Elsevier B.V.All rights reserved.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 2/12
10 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Fig. 1. A schematic diagram of how the vegetation green-up date (GUD) can be
determined using the logistic fittingmethod,modified afterZhang etal. (2003). The
solid line indicatesthe fitted logistic curve,and thedashed line is therate of change
in curvature of the fitted logistic curve. GUD is defined as the first local maximum
of the dashed curve (i.e., point A), and the second local maximum (i.e., point B) is
identifiedas theonset of vegetationmaturity.The redline indicatesthe Local range
(i.e., the range where the vegetation index begin to increase rapidly) in the time-
series data in which GUD is found. (For interpretation of the references to color in
this figure legend, thereader is referred to theweb version of this article)
vegetation indices. The first type is threshold-based, and defines
spring phenology based on the date when a vegetation index
reaches a predefined threshold (i.e., an absolute threshold; Lloyd,
1990) or a specific percentage (e.g.,20% or50%) of itsannualampli-
tude (i.e., a relative threshold; White et al., 1997; Yu et al., 2010).
The second type of method commonly used focuses on changing
characteristics in the time series. This group of methods assumes
that vegetation growth follows a relatively well-defined temporal
pattern and can be fitted by a predefined mathematical function,
which isnormally considered tobea sigmoid function. Springphe-
nologicaleventscan then be identifiedfrom thefittedcurve (Fisher
and Mustard, 2007; Zhang et al., 2003, 2006). For example, Zhang
et al. (2003) used a four-parameter logisticmodel to simulate veg-
etation growthanddefined springphenologyas thedate when the
rate of change in curvature for the fitted curve exhibits the firstlocal maximum (Fig. 1).
All thesemethods, in general, adopt different rules todefine the
area-averagedgreennessonsetin spring, andthusremotely-sensed
springphenologyestimatedbydifferentmethods could differ con-
siderably (White et al., 2009). Zhang’s logistic method actually
captures the timingwhen vegetation greenness begins to increase
rapidly,which is represented by the green-up date (GUD) in Fig. 1.
This method has been widely used in regional and global pheno-
logy research (e.g., Shen et al., 2012; Zhang et al., 2004, 2006; Zhu
et al., 2012), and is currently being used for generating the global
phenological product based on the time series of the Enhanced
Vegetation Index (EVI) data provided by theModerate Resolution
Imaging Spectroradiometer (MODIS; Friedl et al., 2010; Ganguly
et al., 2010; MCD12Q2 User Guide; Zhang et al., 2003, 2006).The effectiveness of logistic methods is dependent on the
basic assumption that vegetation growth follows a well-defined
S-shaped temporal profile. In this study, we hypothesized that
determiningGUDfrom time seriesof a greennessvegetationindex,
especially in grasslands, would suffer from uncertainties in logis-
tic curve fitting due to the fact that the time series from spring to
summer does not necessarily follow an ideal sigmoid curve, and
sometimesmay deviate greatly from this curve. Violation from the
ideal S-shaped growth curve occurs quite often, because natural
vegetation does not grow under ideal conditions, but is instead
affected by a range of environmental stresses (e.g., climate, insects
or diseases)at various times. In temperategrasslands, forexample,
grass growth can be greatly interfered by drought events, because
herbaceous plants usually have underdeveloped root system
Fig. 2. Time seriesfor MODIS enhanced vegetation index (EVI)and thedetermined
GUD (the first localmaximum in the rate of change in curvature) in 2007 (A) and
thecorresponding temporal precipitation in 2007 (B). TheMODISEVI time seriesin
2003 and2005 (C), andthe fitted logistic curve(solid line) andthe determined GUD
in 2003 and 2005 (D). Panel (E) shows the mean air temperature and cumulative
precipitation during thegrowingseason (from Juneto August) and in thepreseason
period (fromNovember of the previous year through April of the current year) for
2003 and2005.Note: Alldata representspatially average data for100×100MODIS
pixels around the Xilinhot weather station, Xilin Gol, Inner Mongolia (43.57◦N,
116.07◦E),which is indicatedas thedashed line boxin Fig. 5. Source of climate data:
ChinaMeteorologicalData SharingServiceSystem(CMDSSS,http://cdc.cma.gov.cn).
comparedwith shrubs and forests and are less resistant to the lackof available soil water (Liu et al., 2013; Zhou et al., 2013). Fig. 2A
illustrates this problem: MODIS EVI time-series data collected in
the Xilin Gol grassland exhibited an obvious two-stage greenness
increase inthe springof 2007.Vegetationgrowth stalled inthemid-
dleof this season(fromapproximatelydayofyear (DOY)140 toDOY
180) due to a lack of precipitation during this period (Fig. 2B). Fit-
ting this time series by Zhang’s logisticmethod (hereafter referred
to as the traditional logistic fitting) only tends tomodel theoverall
growth pattern but miss the “local range” (i.e., the range of data
points that occur around the onset of EVI increase) in the time
series (Fig. 2A). Effects of non-ideal logistic vegetation growth on
GUDdeterminationare further illustrated in Fig.2C andD, inwhich
the difference in the determined GUD is asmuch as 23d between
2003 and 2005. For the two years, an almost identical trajectory
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 3/12
R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20 11
in the EVI time series in the local range suggests that greenness
onsetwasnearlyconcurrent (Fig. 2C). In this example, theproblem
arises from precipitation-induced variations in greenness during
the later period of vegetation growth (from Jun to Aug; Fig. 2E).
Such summerEVIdifference strongly affects the traditional logistic
fitting and leads to a large difference in the GUD determined from
this fitting. Therefore, we shouldpaymore attention to analyze the
local range in the time series inwhich GUD is found.
In this study,we developedan improved logisticmethod,which
we refer to as adaptive local iterative logistic fitting (ALILF), to
improve estimates of GUD from MODIS EVI time-series data. The
ALILF method can address violations of the assumption of sim-
ple S-shaped growth in natural environments, and can determine
GUDmore accurately at the date of the onset of increase inMODIS
EVI. We first present the ALILF method and then provide inter-
comparisons between ALILF and the traditional logistic fitting
method.We didnot evaluate ALILF GUDwith groundobservations
of spring phenological metrics (e.g., bud break or leaf expansion)
because of their different phenological definitions and the incom-
patible spatial scale (i.e., the few plant individuals vs. a pixel
standing for a large area).
2. Methodology
Different from the traditional logistic fitting method, the ALILF
methodmodifies the fitting procedure to closely simulate the tra-
jectory of the MODIS EVI time series in the local range around the
onset ofEVI increase (Fig. 1). Thenewmethodemploys an iterative
technique and an adaptive window and is able to estimate GUD
fromboth logistic andnon-logistic time series. There arefourmain
steps to implement the ALILF method (Fig. 3). We show details of
each step in the Sections 2.1–2.4 and provide the pseudo-code of
the new method in the Appendix A.
2.1. Step1: pre-processing of the MODIS EVI time series
We collected the raw MODIS data, which were used to pro-
duce the Collection 5 MODIS phenological product (MCD12Q2),from the website of the National Aeronautics and Space Admin-
istration (http://reverb.echo.nasa.gov/reverb). First, we generated
raw time-series EVI values from the composited 8-day nadir
bidirectional reflectance distribution function (BRDF)-adjusted
reflectance 500m resolution data (MCD43A4; Schaaf et al., 2002).
We then employed the BRDF-albedo quality product (MCD43A2)
to remove EVI values contaminated by the presence of snowor ice
from the raw time series. Assuming that vegetation is biologically
inactive at low temperatures, we used the land surface tempera-
ture product (MOD11A2; Wan et al., 2002) to remove winter EVI
values forwhich thetemperaturewaslower than 0 ◦C. This process
can account for cases in which the data quality label (MCD43A2)
fails to identify snowor ice.Wefilled the resulting gaps in the time
series by linear interpolation,anduseda three-pointmedian-valuefilter tosmooththe timeseries (MCD12Q2User Guide;Zhanget al.,
2006).
2.2. Step2: modeling the general pattern of the MODIS EVI time
series
Although vegetation growth in natural environments does not
follow an ideal sigmoid curve, it nonetheless exhibits a general
growth pattern in which the vegetation starts to grow in early
spring and reaches its full bloom in summer (Fig. 2). Thus, we used
a logistic model to simulate this general pattern with the goal of
obtaining somepreliminary information about vegetationgrowth.
To simulate vegetation growth, we first determined the time
period in which EVI exhibited a sustained increase, and used this
Fig. 3. Flowchart for the adaptive local iterative logistic fitting (ALILF) method.
NBAR: nadir bidirectional reflectance distribution function (BRDF)-adjusted
reflectance.
period for curve fitting (Ganguly et al., 2010). In practice, the tran-
sition forEVI from increasing todecreasing trendwas identifiedby
a change from positive to negative linear slopeusing thefive-point
movingwindow.Twocriteriawere further adopted toexclude spu-
rious transitions in the EVI time series: the maximum value in the
identified period should exceed 70% of the annual maximum EVI,
and the change in EVI within the identified period exceeded 35%
of the annual amplitude of the change in EVI. We then applied
Zhang’s four-parameter logistic equation to the EVI time series in
the identified period (Zhang et al., 2003):
EVI(t ) =c − d
1+ exp(a+ bt ) + d (1)
where t is the DOY. The parameter c in Eq. (1) is slightly modified
compared to that formulated by Zhang et al. (2003). In the present
equation, c and d indicate the maximum and background EVI val-
ues, respectively, and a and b control the shape of the curve. In the
non-linear least-square curve fitting, we constrained the parame-
ter c to be EVITP (EVI at the transition point, Fig. 4A) plus orminus
10% (i.e., EVITP±0.1EVITP, which we will discuss in Section 4.1).
We did not constrain parameter d in this fitting because of prob-
able contamination of the background EVI observations by noise.
Through the constrained non-linear curve fitting, we first gained
thefitting valuesfor theparameters c and d, hereafterreferredto as
c f and df , which characterizetheannualamplitudeof thechange in
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 4/12
12 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Fig.4. A schematicdiagramshowing (A)determinationof the localwindowand(B)
the two criteria (cases 1 and 2) for convergence for ALILF, see Sections 2.3 and 2.4
for details.
vegetationgreenness.We thenestimatedGUDfromthefittedcurve
basedonthefirst localmaximum intherateof changein curvature,
andthisGUDestimate isdeemedas theinitialGUD. Theobtainedc f ,
df and the initial GUD provide the basis for refining GUD estimates
in subsequent steps.
2.3. Step3: iterative logistic fitting in the local range
We developed an iterative logistic fitting technique to further
adjust thevegetation growthsimulation. This methodensures that
the fitted curve captures the annual amplitude of the change in
greennesswhile also providing a better simulation of thepoints in
the local range where EVI begins to increase rapidly.
Specifically, we first defined a local window consisting of
nine consecutive EVIobservations (approximately a 2-month time
span), with four points on either side of the initial GUD (Fig. 4A).
We then used Eq. (1) to fit only EVI data in the local window by
implementing a constrained non-linear optimization algorithm. In
the fitting, parameter c was fixed at the value of c f and parameter
dwas constrained to bewithin the range of df ±0.2(c f −df ). In this
manner, a new GUD is determined from the fitted logistic functionfor the local range. Based on the newly determined GUD, a new
nine-point local window can be defined, and Eq. (1) is used again
to fit EVI data in the new window. This process is repeated. Dur-
ingthis iterativeprocess, it probably encounters theexceptionthat
there are less than four points on one side of the determined GUD
when GUD is determined approximately at either end of the iden-
tified timeperiod. For this case, we require that at least two points
should be found on either sideof the determinedGUD for the local
fitting (i.e., at least seven points in the local window). Otherwise,
we employed all points within the identified time period for the
fitting.
The iterativeprocess is continuing, andgradually, thetrajectory
of the EVI time series in the local window tends to become consis-
tent with the fitted logistic curve around the position of max rate
of change in curvature. In other words, the determined GUD con-
verges on the date of the onset of the increase in EVI. Two possible
conditions indicate that the iteration should end: the determined
GUD stops changing between consecutive iterations (Fig. 4B, case
1);or thesameGUDbeginto reappearafter several iterations, lead-
ing to oscillation between the two values with a certain period
(Fig. 4B, case 2). In practice, the iteration is convergent because
there is a finite total numberof differentnine-point localwindows
for a given growing season. We discuss the influence of the ini-
tial GUD on the iteration as well as the effect of the parameter
constraint further in Section 4.1.
2.4. Step4: determining the final GUD from an adaptive local
window
For the two cases of convergence, we adopted different strate-
gies to determine the final GUD.
When the iteration converges on a final value, we defined the
final GUD as the convergence value, which depends on the fitting
for ninepointsin the localwindow(Fig.4B, case1). This caseis nor-
mally encountered when there is an obvious increase in EVI in the
local range. However, when there is a less obvious increase in EVI
within the local rangedue tonoisy data, this can induce oscillationbetweentwoGUDvaluesduringtheiteration.Inthiscase,wedefine
an adaptive local window that includes the four points below the
minimum GUD, the four points above the maximum GUD, and the
pointsin theoscillation(Fig.4B, case2). ThefinalGUDisthusdeter-
mined by fitting EVI points in this larger window. How to end the
iteration depends on the quality of the EVI time-series data in the
local range. The ALILF method can flexibly adjust the width of the
local windowusing this adaptive technique and is therefore resis-
tant to a range of noise levels, as we will demonstrate in Section
4.1.
Before exporting thefinal GUD, we used an additional criterion
to filter theresults (the decisionpoint labeled “condition” in step 4
of Fig. 3). We assumed that the time interval between GUDand the
timingof vegetationmaturity (i.e.,the time periodbetweenA andBshown inFig. 1) shouldbe at least 1month(i.e., approximately four
EVI observations). If this condition was not met, we employed the
initial GUD as the final output. This constraint was used to ensure
that if the final GUD was determined by fitting the local range of
time series, EVI points after the maturity date cannot be included
in thefitting. Forsiteswith anunusuallyshortgrowing season(e.g.,
alpine desert), this constraint maybemore frequently met.
2.5. Application
We evaluated the ALILF method by applying it to estimate GUD
in the Xilin Gol grassland from MODIS EVI time-series data dur-
ing 2002–2010. The Xilin Gol region is located in Inner Mongolia,
China, and covers an area of 230,000km2, stretching from 41.5◦Nto 46.9◦N and from 111.2◦E to 119.9◦E (Fig. 5). In this region, the
annualmeanair temperature is2.4◦C andtheannualmeanprecipi-
tationrangesfrom250 to350mm,mostofwhichfallsbetweenMay
andSeptember(Zhuoet al., 2007). Precipitation showsawestward
decreasing trend in the spatial distribution, ranging from above
400mm in the eastern region to below 150mm in the western
region. Grasslands dominate Xilin Gol with steppes andmeadows
distributed in the central and eastern areas and desert steppes in
the western region, as is indicated by the 1:1,000,000 vegetation
map of China (Environmental and Ecological Science Data Center
for West China, http://westdc.westgis.ac.cn). Spring phenology of
Xilin Gol grassland was controlled by both temperature and pre-
cipitation (Yu et al., 2003; Liu et al., 2013) and water availability
was shown to have stronger influences (Liu et al., 2013).
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 5/12
R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20 13
Fig. 5. Location of the Xilin Gol grassland of China’s Inner Mongolia Autonomous
Region, and distribution of the main vegetation types in Xilin Gol. Source of the
vegetation data: http://westdc.westgis.ac.cn(inChinese), see thetext fordetails.
Before applying the ALILF method, we followed Shen et al.
(2014) and adopted three criteria to exclude some EVI time-series
data in this region due to the lack of vegetation and seasonality.
ALILF was not implemented if any of the following situations was
encountered: (1) the annualmaximum of EVI did not occur within
June–September; (2) the average EVI from July to September was
smaller than0.08,and(3) this averagewassmaller than1.2times of
the background EVI (i.e., the average EVI beforeApril of each year).
3. Results
Fig.6 presents theperformanceoftheALILFmethod forthethree
EVItimeseriesshowninFig.2. Ingeneral,thefittedcurvesshoweda
good ability to simulate thechanging characteristicsof EVIfor localrange inall three years. The GUDwere found tobedetermined reli-
ably at thedateof theonsetofvegetation greennessafterourvisual
inspection of thegraphs (dotted blue lines in Fig. 6A–C). Thediffer-
ence inGUD between 2003and 2005decreased from23d (Fig. 2D)
to 10 d (Fig. 6A and B), which suggests that the ALILF method
was able to account for the influence of inter-annual variations in
greenness on theGUDdetermination. Furthermore, unlike the tra-
ditional logisticmethod, theALILFmethodwas able to address the
precipitation-induced two-stage vegetation growth curve in 2007
(Fig. 6C). To obtain the final GUD, we performed curve-fitting four
times in a nine-point window for the 2003 time series and found
that GUD converged on a fixed value. For EVI curves in 2005 and
2007, however, the iteration showed oscillation without conver-
gence, so the width of the fittingwindowwas adaptively adjustedto include 10 EVI observations.
Fig. 7 compares the regional distribution of GUD determined
using the ALILF method and the corresponding MODIS phenology
product (MCD12Q2). A general spatial trend of increasing GUD
moving from southeast to north and northwest can be seen for
both ALILF and the MCD12Q2 product, which is geographically
consistent with known spring phenological events in theXilin Gol
grassland (Li et al., 2013; Wang et al., 2006). This spatial pattern of
GUD, however, also exhibited some differences between years. It
generally had later GUD in 2006, which was probably due to the
low temperature and the utter lack of precipitation before DOY
120 (Figs. S1 and S2). GUD in 2009 showed less obvious westward
increasing trend and vegetation turned green almost before May,
which might be explained by the abundant precipitation during
March-April across the entire region (Fig. S2). GUD in 2010 were
mainlytakenplacefromlatelyAprilto earlyMay,and thiswasprob-
ably causedby theextremely lowtemperatureduringMarch–April
(Fig. S2)and thegreatest increase in temperature fromApril toMay
for this year (Fig. S3).
Theabsolute difference between ALILFGUDandMCD12Q2 (i.e.,
|difference| in Fig. 7) were usually within 1 month, and there was
no specific spatial pattern for this difference. The largest GUD dif-
ferences occurred in thewest, central part or northeastof XilinGol
for different years. We identified two sub-regions with high val-
ues of |difference| in 2002 and 2007 and further analyzed them
in Figs. 8 and 9, respectively. The comparisons for the sub region
of 2002 showed that ALILF GUD was more continuous in terms of
the spatial distribution; ButMCD12Q2 was quite fragmented with
abrupt changes of GUD within short spatial distances (Fig. 8 A and
B), which is impossible for phenology of natural vegetations. We
additionally visually checked the land cover of this sub region for
the same year by a high-spatial-resolution satellite image (Landsat
TM), and excluded the possibility of cropland in this sub region.
So the discontinuous spatial distributionofMCD12Q2 appeared to
be inconsistent with the reality. The histogram of GUD revealed
the difference between the two methods. A very small number of
pixels turned green before April estimated by the ALILF method,
whereas more than 20% of pixels had GUD inMarch for MCD12Q2
(Fig. 8C).We further calculated the average EVI time series aswell
as theaverage GUDplus thestandarddeviations, for allpixelswith
GUD estimations within the range of DOY [60,90] fromMCD12Q2
but outside this range when using the ALILF method (Fig. 8D). The
investigations showed that ALILF GUD matched the date of the
onset of greenness increase better than MCD12Q2. Similar analy-
ses for another sub region located in the central part of Xilin Gol in
2007 also indicatedtheabilityofALILFtocapturetherealgreenness
onset in regional applications (Fig. 9).
We present the performances of the ALILF method and the tra-
ditional logistic fitting for some representative EVI time series. It
includesa serieswithawell-definedEVItemporal profile(Fig. 10A),
a series with background EVI contamination due to snow and ice
(Fig. 10B), several series with one or more disturbances of EVI inthelatergrowth periodafterGUD(Fig. 10C–E), andserieswith poor
data quality in the local range (Fig. 10F and G). The ALILF method
bettersimulated thetrajectoryof time-seriesdata inthelocalrange
and captured greenness onset effectively in all cases. In contrast,
performancesof the traditional logistic fittingwere unstable and it
lost efficiency in some caseswith non-logisticgrowthornoise con-
taminations. The newmethod appeared to demonstrate an ability
to resist data noise of EVI time series. For instance, for the casesof
noisy time series with a less obvious local range in Fig. 10F and G,
the ALILF method adaptively enlarged the windowwidth and was
also able to estimate a reasonable GUD.
It is also noteworthy that it seems somewhat arbitrary to iden-
tify thesustainedperiodofincreasingEVIforlogisticfittingby using
thefive-point movingwindow technique (Ganguly et al., 2010). Inthe example of vegetation growth with two stages (Fig. 10H), the
transitionpointwouldprobably be identifiedat thelocalmaximum
during the first stage or at its annual maximum (Fig. 6C). As we
expected, thereis nouniversalempiricalcriteriontoperfectly iden-
tify the lengthof thesingle growing seasonparticularly in regional
applications. The identification errors are thus inevitable for some
pixels andprobablyled touncertainties inGUDestimationbasedon
the traditional logistic fitting. However, such errors seemto affect
ALILF less because of the new fitting principle (Figs. 6C and 10H).
For example, if the transition is identifiedat the annualmaxima in
Fig. 10H, GUDestimatedbytheALILFmethod variedlittle (DOY120
vs. 119; Fig. 11) but the traditional logistic fitting showed a large
variation (DOY 92 vs. 119), suggesting the advantage of the new
method in being able to resist period identification errors.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 6/12
14 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Fig. 6. (A–C) The fitted curves (blue solid curve) and final determined GUD (blue dashed vertical lines) by using the ALILF method to analyze the three EVI time series in
Fig. 2. The right columnshowsthe corresponding process of iterations (blue solid polylines) in using theALILF method. (For interpretation of thereferences to color in this
figure legend, thereader is referred to theweb version of this article)
4. Discussion
4.1. Performances and advantages of the ALILF method
We developed and applied the ALILF method to determine
GUD using data for the Xilin Gol grassland of Inner Mongolia.This method is conceptually improved because it compensates
for an invalid assumption used in the traditional logistic fitting
method. As the examples described in the results section show,
the EVI time series is often far from a simple S-shaped pattern
because environmental factors,mainly temperature andprecipita-
tion,control vegetationgrowth undernatural conditions. Although
the air temperature usually follows an S-shaped temporal profile
(Villegas et al., 2001, Zhang et al., 2003), precipitation events are
often temporally unpredictable, which can greatly affect growth
trajectories in a grassland (Fig. 2). To demonstrate the effects of
precipitation on grass growth, we further investigated the rela-
tionshipbetweenvegetationgreennessduring growing season and
concurrent temperature and precipitation around all nine mete-
orological stations in the Xilin Gol grassland (Fig. S1). We found
that precipitation greatly increase EVI and has much stronger
effects on EVI than temperature (Fig. S4), which suggests that the
vegetation growthcanbe strongly affected or even stall when pre-
cipitation is lacking and thus precipitation-driven grassland such
as Xilin Gol is likely to exhibit non-logistic patterns of growth.
To address deviations from the S-shaped growth curve, the ALILF
method adopts an iterative technique and the use of an adaptive
window. Therefore, ALILF is a more general method that has an
improvedabilityto estimateGUDfromboth timeserieswitha well-
defined S-shaped curve and time serieswith non-logistic patterns
(Figs. 6–10).
Since ALILF is based upon logistic fitting of the local range,
one concern might be whether the method is much more sensi-
tive to noisy EVI data. To address this concern, we conducted a
simulation experiment to quantify the effects of different noise
levels on the method’s performance. To do so, we added random
Gaussian-distributednoise,withamean valueof zero andstandard
deviations ranging from 0% to 15% of the EVI value, to an ideal
logistic curve with a true GUD at DOY 120 (Fig. 12A). We consid-
ered twoscenarios in thesimulation experiment. In scenario 1, thetime seriesoutside the local range isperturbedby noise.Weaim to
examine whether the true GUD can be estimated from these noisy
timeseries, byusingboththe traditional logisticfittingandALILF. In
scenario 2, EVIwithin thelocalrangeis perturbedby noise.Because
a highernoise level indicatesa less obvious changing characteristic
for the local range, this scenario is designed to examine how the
width of local window in the ALILF method is related to the noise
level. To obtain a credible result, we repeated the simulation 100
times at each noise level.
In scenario 1, we calculated the absolute difference ( AD)
between GUDestimations and the true GUD, and found increasing
AD with noise levels for both ALILF and traditional logistic fitting
(Fig. 12B). There was no significant difference for AD from both
methods at low noise levels (P >0.01, two independent samples t -
test),whereas at highernoise levelsof 10%and 15%, AD of theALILF
method was significantly smaller than that of traditional logistic
fitting (P <0.01). It suggests that the new method is not more sen-
sitive to noisy data compared with the traditional logistic fitting
method. In scenario 2, the width of the local window in the ALILF
method is found to increase from 9 tomore than 13 points as the
noise level increases from0% to15%,which shows a significant lin-
ear positive correlation (R2 =0.96, P <0.01; Fig. 12C). This confirms
that the width of local window can be adaptively adjusted accord-
ing to the quality of the EVI time-series data in the local range. This
adaptive technique ensures the practicability of the ALILF method
for noisy time series. Without this technique, it would be difficult
toapply thenewmethodto noisy time-seriesdata, especiallywhen
there is high noise in the local range.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 7/12
R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20 15
Fig. 7. Spatial distributions and statistical histogram of GUD estimations using the ALILF method and those obtained from the MODIS Land Surface Phenology Product
(MCD12Q2) from 2002 to 2010 forthe Xilin Golgrassland of China’s InnerMongolia Autonomous Region. Therightmost columnshows theabsolutedifference (|difference|)
of the two GUDestimations for each year. Twosub regions with large differences in 2002 and2007 areindentified.
Another advantage of the ALILF method is that it is less affected
by period identification errors (Fig. 11). We need to employprede-
fined criteria to identify the timeperiodof increasing EVI for curve
fitting. However, anycriteria aredefinedempirically andcertainly,
they arenotuniversally applicable. TheALILFmethod is tolerant to
this error to someextent because of its fitting in the local range of
time series, which further suggests the robustness of the method
for regional-scale application.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 8/12
16 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Fig. 8. The spatial distribution of GUD for the sub region of 2002 estimated fromMCD12Q2 (A) and from the ALILF method (B). The statistical histogram for the two GUD
estimations (C), and (D) the average EVI time-series data (circle) and the average GUDplus thestandard deviations, for all pixelswith GUDestimationswithin the range of
[60,90] fromMCD12Q2but outside this rangewhen using theALILFmethod.
TheALILFmethodadoptsa constrainednon-linear least-squares
fitting in its implementation (Fig. 3). It is reasonable to constrain
the parameter c in Eq. (1) to obtain an ecologically meaningful
maximum EVI. However, an accurate c value is not necessary in
the ALILF method because the method focuses on the changing
characteristics of EVI in the local range, so its estimates of GUD
are less sensitive to the c value. Therefore, we constrained c to be
within 10% of the EVI value at the transition point in step 2, and
fixed the estimated c f in step 3. In a similar manner, the ALILF
method constrains parameter d to be df (the fitted initial value
of d produced in step 2), plus a variation of 20% in (c f −df ) inthe fitting. This guarantees that the curve-fitting from the ALILF
method captures the annual amplitude of the change in greenness
while also accounting for some noise values in thebackground EVI
(Fig. 10B). However, the ALILF method does not make efforts to
improve thepre-processingonnoiseEVI, such as thesnow/ice con-
tamination in thebackgroundEVI andcloudcontaminationleading
to gaps in the time-series data. Although a number of methods
have been developed to filter noise and to reconstruct vegeta-
tion indices (e.g., Jönsson and Eklundh, 2002; Chen et al., 2004),
we used the simplest linear interpolation to fill gaps in the EVI
time series. Our investigations showed that in grasslands, tempo-
ral profile of EVI could inherently exhibit various shapes including
the two-stage increase in greenness. So we should take caution
when using a predefined filter (e.g., Gaussian filter; Jönsson and
Eklundh, 2002) to construct time series data. Without a prioriknowledge of vegetation growth, we recommend using the lin-
ear interpolation to fill gaps. Nevertheless, large errors could be
expectedoncenumbersofconsecutiveEVIaremissing (Zhangetal.,
2009).
Fig. 9. (A–C) Similar to Fig.8 butfor thesub regionof 2007. (D)The average EVItime-series data (circle) andthe average GUDplus thestandard deviations, forall pixelswith
GUD estimationswithinthe range of [60,75] fromMCD12Q2 butoutside this rangewhen using theALILFmethod.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 9/12
R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20 17
Fig. 10. Applying theALILFmethodand thetraditional logistic fitting to some representativeMODIS EVI time series to estimate GUD. Each panel includes an enlargedview
of the local range aroundthe determined GUD. WW:thewidthof local windowused by ALILF.
Finally, we suggest that the accuracy of GUD estimation using
ALILF would be less affected by the initial GUD because of the iter-
ative procedure and the parameter constraints used in the fitting.To support this belief, we conducted a simulation experiment.We
first assumed that the initial GUD varied within 1 month before
or after the final determined GUD for the three EVI time series
in Fig. 6. We then used the ALILF method to estimate GUD. The
results confirm that the final determined GUD is independent of
the initial GUD despite the use of different numbers of iterations
(Fig. 13). Thenumberof iterations usually increases (e.g. from four
to eight in 2007), with an increasing difference between the initial
and final determined GUD. Therefore, a more accurate initial GUD
could improve thecomputational efficiency. Atpresent,we recom-
mendgenerating the initial GUD in the way described in this paper
for estimation of regional GUD using the ALILF method.
4.2. Applicability of the ALILF method
We used the ALILF method to analyze MODIS EVI time-series
data. This source of data is considered tobeof highquality because
theeffects of clouds, viewingangles, andatmosphericaerosolshave
been greatly minimized (Huete et al., 2002; Schaaf et al., 2002). In
practice, the ALILFmethod canbe also applicable to other satellite
time series, such as the longer recordof thenormalized-difference
vegetation index (NDVI) time series provided by the Advanced
VeryHighResolutionRadiometer (AVHRR).However,AVHRRNDVI
time series may suffer from low data quality due to the fact that
the AVHRR sensors were not originally designed for vegetation
monitoring and their near-infrared bands cover an atmospheric
water vapor absorption feature (Cihlar et al., 2001; Yuet al., 2013).
Recent studies found that low dataquality of AVHRR NDVI leads to
Fig. 11. Applying the ALILF method and the traditional logistic fitting to the time
series in Fig. 10Hwhen thefitting periodis determined to be from thebeginning to
theannual maximum EVI.
detection of a false decadal trend in GUD in the Tibetan grassland
(Yu et al., 2010; Zhang et al., 2013). Therefore, we recommend
applying the ALILF method to the more advanced MODIS EVI
data and potentially its heritage provided by Visible/Infrared
Imager/Radiometer Suit (VIIRS) instrument onboard the S-NPP
satellite (Vargas et al., 2013), which enable monitoring of vege-
tationphenology continuously at the global scale.
Although MODIS EVI products have improved the spatial res-
olution to 500m, a MODIS pixel potentially represents a mixture
of plant species that might have substantially different phenology,
and this variation poses a challenge to defining a representative
GUDwithinagivenpixel.A fewstudieshavesuggestedthatin some
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 10/12
18 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Fig.12. (A)An ideal logistic curve(thefittedcurve in Fig. 6A) with a true GUDat DOY120 used forthenoisesimulationexperiment.Noise wasapplied tothe time-seriesdata
outside the localrange (scenario 1) or only to the local range (scenario 2). (B) In scenario 1, the absolute error between GUD estimation and the trueGUD (themean value
plus the standard deviation for 100 random iterations) at different noise levels. * indicates a significant difference of the absolute error between ALILF and the traditional
logistic fitting (P <0.01, two independent samples t -test). (C) In the scenario 2, the relationship between the width of the local window used in the ALILF method and the
noise level.
forests, theremight bea large phenologicaldifference between the
understory shrubs and grasses, and the overstory canopy (Badeck
et al., 2004; Richardson and O’Keefe, 2009), and this can confuse
remotelysensedsignalsandaffectGUDdeterminations.At present,
none of theexistingmethods (including ALILF) cansolve this issue.
For such complex scenes, it would be more meaningful to find
ways to discriminate among different spring phenological events
for different species by decomposing the time-series data. Such ananalysis is beyond the scope of this study, and will be explored in
our future phenological research.
Our study highlighted the importance of carefully considering
characteristics of vegetation growth when estimating phenology
from time-series data by logistic fitting. Several recent studies also
noticed this issue andhave improved simulations of time-series of
vegetation indices to better estimate phenology (Che et al., 2014;
Elmore et al., 2012). As satellite-derived vegetation phenology has
been increasinglyused by theresearchcommunities of globalecol-
ogyandclimate change,wecall formoreattentiontothe phenology
detection for various ecosystems to provide reliable data basis for
ecological applications.
4.3. Conclusions
Non-ideal growth trajectories are probably most common in
certain typesof ecosystems, particularlygrasslands.The traditional
logistic fitting method actually does not account for these varia-
tions in that it does not model the EVI trajectories accurately. The
ALILF method, however, appears to yield GUD dates that are less
Fig. 13. ThedeterminedGUDandnumber of iterations when different initial GUDvalues areused in theALILFmethod forthe three EVI time seriesin Fig.6. The initial GUD
values were assumed to vary within1 month before or after thefinal determined GUD. Thedottedarrowsindicate theinitial GUDthat is generatedby theALILFmethod.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 11/12
R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20 19
influenced by the variations and are closer to the actual time that
the land surface began to green up. The new method is resistant
to a range of noise levels because of an adaptive window, which
enable ALILF to enlarge the windowwidth used for logistic fitting
when noise level increases. In addition, the ALILF method is less
affected by the definition of the maximum EVI and the identifica-
tion of thefittingperiod. These advantages suggest that this simple
newmethodmaybewidely used formonitoring springvegetation
phenologybased on satellite time-series data.
Acknowledgments
We thank three anonymous reviewers and the associate editor
Wagner-Riddle for their detailed and constructive comments that
helpedus to improve themanuscript. This studywas supportedby
the project of early detection and prediction of climate warming
based on the long-term monitoring of fragile ecosystems in East
Asia funded by the Ministry of Environment, Japan, and a grant to
M.S. from the National Natural Science Foundation of China No.
41201459.
Appendix A. Appendix A:
The followingpseudo-code describes the ALILFmethod
// Input parameters:
evi= time series data with time resolution of 8-d
ts,tp = thesequential number for the start (land surface temperature> 0) and
transition of the evi time-series, respectively
// Outputs:
gud=green-up onset date
// the equal-weighted logistic fitting for the first estimate
wts[all]← 1.0
gud=Getgud (evi, wts, 1)
GUD result[1]← gud
// the iterative logistic fitting
iternum=2
WHILE (iternum≥ 2) {
centerpoint=Round(( gud-1)/8) +1 // determine thecenter point of thewindow
wts=Getwts (centerpoint, centerpoint )
gud=Getgud(evi, wts, iternum) // call the functions
GUD result[iternum]← gud
FOR i =1 to iternum-1 {
IF (GUD result[i]= GUD result[iternum]) THEN {
IF i = 1 THEN {
gud← GUD result[1]
iternum← 0
Break
ELSE
maxgud←Max (GUD result[i to iternum])
mingud←Min (GUD result[i to iternum])
maxpoint ← Round ((maxgud -1)/8)+1
minpoint ← Round ((mingud-1)/8)+1
wts=Getwts (minpoint, maxpoint )
gud=Getgud (evi, wts, iternum)
gud= (gud-maturity date>30) ? gud: GUD result[1] // the additional
// constraint in step 4; thematurity date is calculated based on
// thefour fitting parameters
iternum← 0Break
} //end if
} //endif
} // end for
Iternum+ +
}// endwhile
//two functions
Functionwts=Getwts(centerpoint, centerpoint )
Global tstp //global variables
// theintersection of thetwo vectors
intersect= [centerpoint -4 to centerpoint+ 4]∩[ts to tp]
IF Length(intersect ) ≥ 7 THEN {
// at least seven pointsare included in thefittingduring iteration
wts[all]← 0.0
wts[intersect]← 1.0
ELSE
Appendix A: (Continued)
wts[all]← 1.0
}
Function gud=Getgud (evi, wts, iternum)
Global tp c f df // global variables
IF iternum= 1 THEN {
Limits c : evi[tp]±10%evi[tp]
[c f , df , gud]= Logistic fitting (evi, wts,c )
ELSE
Fixes c : c f AND Limitsd: df ±20%(c f −df )gud= Logistic fitting (evi,wts, c,d) // forthe non-linear least-square
// fitting procedure, refer to http://purl.com/net/mptif
}
Appendix B. Supplementary data
Supplementary material related to this article can be found,
in the online version, at http://dx.doi.org/10.1016/j.agrformet.
2014.09.009.
References
Badeck, F.W., Bondeau, A., Böttcher, K., et al., 2004. Responses of spring phenology
to climate change. NewPhytol. 162, 295–309.ChapinIII, F.S., Randerson, J.T.,McGuire, A.D., et al., 2008. Changingfeedbacksin theclimate-biosphere system. Front. Ecol. Environ. 6 (6), 313–320.
Che,M.L., Chen, B.Z., Innes, J.L.,et al.,2014. Spatial andtemporalvariations in theenddate of the vegetation growing season throughout the Qinghai-TibetanPlateaufrom 1982 to 2011. Agric. For. Meteorol. 189–190, 1711–1722.
Chen, J., Jönsson, P., Tamura, M., et al., 2004. A simplemethod for reconstructing ahigh-qualityNDVI time-series dataset basedon theSavizky–Golayfiter.RemoteSens. Environ. 91, 332–344.
Cihlar, J., Tcherednichenko, I., Latifovic, R., et al., 2001. Impact of variable atmo-spheric water vapor content on AVHRR data corrections over land. IEEE Trans.Geosci. RemoteSens.39, 173–180.
Cong, N.,Piao, S.L., Chen, A.P., et al., 2012. Spring vegetation green-update in Chinainferred from SPOT NDVI data: a multiple model analysis. Agric. For. Meteorol.165, 104–113.
Elmore, A.J., Guinn, S.M.,Minsley, B.J., et al., 2012. Landscape controls on the timingof spring, autumn, and growing season length in mid-Atlantic forests. GlobalChange Biol. 18, 656–674.
Fisher, J., Mustard, J., 2007. Cross-scalar satellite phenology from ground, Landsat,
and MODIS data. Remote Sens. Environ. 109, 261–273.Friedl, M.A., Sulla-Menashe, D., Tan, B., et al., 2010. MODIS Collection 5 global land
cover: algorithm refinements and characterization of new datasets. RemoteSens. Environ. 114, 168–182.
Ganguly, S., Friedl, M.A., Tan, B., et al., 2010. Land surface phenology from MODIS:characterization of theCollection5 global landcover dynamicsproduct.RemoteSens. Environ. 114, 1805–1816.
Huete,A., Didan,K., Miura,T., et al., 2002.Overviewof the radiometric andbiophys-ical performance of the MODIS vegetation indices. Remote Sens. Environ. 83,195–213.
Jeong, S.J., Ho, C.H., Jeong, J.H., 2009. Increase in vegetation greenness and decreasein springwarming over east Asia. Geophys. Res. Lett. 36, L02710.
Jeong,S.J., Medvigy,D., Shevliakova,E., et al., 2012.Uncertaintiesin terrestrialcarbonbudgets related to spring phenology. J. Geophys. Res.: Biogeosci. 117, G01030,http://dx.doi.org/10.1029/2011jg001868.
Jönsson,P., Eklundh,L., 2002. Seasonalityextractionbyfunctionfittingto time-seriesof satellite sensordata. IEEE Trans.Geosci. RemoteSens.40, 1824–1832.
Li,X.H., Chen, S.H.,Han,F., 2013.Effectsof droughton grasslandturning greenperiodin Inner Mongolia. Pratacultural Sci. 30 (3), 452–456 (In Chinese with abstractin English).
Liu, H., Tian, F., Hu, H.C., et al., 2013. Soil moisture controls on patterns of grassgreen-up in Inner Mongolia: an index based approach. Hydrol. Earth Syst. Sci.17, 805–815.
Lloyd, D., 1990. A phenological classification of terrestrial vegetation cover usingshortwavevegetationindex image.Int. J. Remote Sens. 11 (12), 2269–2279.
MCD12Q2 User Guide. http://www.bu.edu/lcsc/files/2012/08/MCD12Q2UserGuide.pdf (Last updated: 8 August2012).
Menzel, A., Sparks, T.H., Estrella, N., et al., 2006. European phenological responseto climate change matches the warming pattern. Global Change Biol. 12 (10),1969–1976.
Partanen,J., Koski, V.,Hänninen,H.,1998. EffectsofphotoperiodandtemperatureonthetimingofbudburstinNorwayspruce(Piceaabies). TreePhysiol. 18,811–816.
Piao,S.L.,Fang,J.Y.,Zhou, L.M., et al., 2006. Variations in satellite-derived phenologyin China’s temperatevegetation.Glob. Change Biol. 12 (4), 672–685.
Piao,S.L.,Ciais,P., Friedlingstein, P.,et al.,2008.Netcarbondioxidelossesof northernecosystems in response to autumnwarming. Nature451, 49–52.
Richardson, A.D.,Bailey,A.S., Denny, E.G.,et al., 2006.Phenology of a northernhard-
wood forestcanopy. Glob. Change Biol. 12 (7), 1174–1188.
8/10/2019 An improved logistic method for detecting spring vegetation.pdf
http://slidepdf.com/reader/full/an-improved-logistic-method-for-detecting-spring-vegetationpdf 12/12
20 R. Cao et al./ Agricultural andForestMeteorology 200 (2015) 9–20
Richardson, A.D.,Hollinger,D.Y., Dail,D.B., et al.,2009. Influenceof springphenologyon seasonal andannual carbonbalance in twocontrasting NewEngland forests.Tree Physiol. 29, 321–331.
Richardson, A.D.,O’Keefe, J., 2009.Phenologicaldifferences betweenunderstoryandoverstory:a casestudyusingthelong termHarvardForestrecords.In:Noormets,A. (Ed.), Phenology of Ecosystem Processes. Springer, New York, pp. 87–117.
Schaaf, C.G., Gao, F., Strahler, A.H., et al., 2002. First operational BRDF, albedonadirreflectance products fromMODIS.Remote Sens. Environ. 83, 135–148.
Shen, M.G., Tang, Y.H., Chen, J., et al., 2011. Influence of temperature and precip-itation before the growing season on spring phenology in grasslands of the
central and eastern Qinghai-Tibetan Plateau. Agric. For. Meteorol. 151, 1711–1722.
Shen,M.G., Tang, Y.H., Chen, J.,et al., 2012. Specification of thermal growing seasonin temperateChina from 1960 to 2009. Clim. Change 114, 783–798.
Shen,M.G.,Zhang,G.X.,Cong,N., et al.,2014. Increasingaltitudinalgradient of springvegetation phenology during the last decade on the Qinghai-Tibetan Plateau.Agric. For. Meteorol 180–190, 71–80.
Villegas,D., Aparicio,N., Blanco, R., Royo, C.,2001. Biomass accumulation andmainstemelongationof durum wheat grown under Mediterranean conditions. Ann.Bot. 88, 617–627.
Wan, Z.M., Zhang, Y.L., Zhang, Q.C., et al., 2002. Validation of the land-surfacetemperature products retrieved from Terra Moderate Resolution Image Spec-troradiometer data. Remote Sens. Environ. 83, 163–180.
Wang, H., Li, X., Yu, H., 2006. Monitoring growing season of typical steppe innorthern China based on NOAA/AVHRR NDVI data. J . Plant Ecol. 30, 365–374.
White, M.A., Thornton, P.E., Running, S.W., 1997. A continental phenology modelfor monitoring vegetation responses to interannual climatic variability. Glob.Biogeochem. Cycles 11, 217–234.
White, M.A., de Beurs, K.M., Didan, K., et al., 2009. Interconparison, interpretation,and assessment of spring phenology in North America estimated from remotesensing for 1982–2006. Glob. Change Biol. 15, 2335–2359.
Vargas, M., Miura, T., Shabanov,N., et al., 2013. An initial assessment of SuomiNPPVIIRS vegetation index EDR. J. Geophys. Res. Atmos.118, 12,301–12,306.
Yu,F.F., Price,K.P., Ellis,J., etal., 2003. Responseof seasonal vegetation developmentto climatic variations in eastern central Asia. RemoteSens.Environ.87, 42–54.
Yu, H.Y., Luedeling, E., Xu, J.C., 2010. Winter and springwarming result in delayedspring phenology on the Tibetan Plateau. Proc. Natl. Acad. Sci. USA 107 (51),22151–22156.
Yu, Z., Wang, J.X., Liu, S.R., et al., 2013. Inconsistent NDVI trends fromAVHRR, MODIS, SPOT sensors in the Tibetan Plateau. In: Proceedings of the Second International Conference on Agro-Geoinformatics, pp. 97–101,http://dx.doi.org/10.1109/Argo-Geoinformatics.2013.6621887.
Zhang,G.L., Zhang, Y.J.,Dong,J.W., etal., 2013.Green-updates in theTibetanPlateauhave continuously advanced from 1982 to 2011. Proc. Natl. Acad. Sci. USA 110(11), 4309–4314.
Zhang, X.Y., Friedl, M.A., Schaaf, C.B., et al., 2003. Monitoring vegetation phenologyusingMODIS.Remote Sens. Environ. 84, 471–475.
Zhang, X.Y., Friedl, M.A., Schaaf, C.B., 2006. Global vegetation phenology fromModerate Resolution Imaging Spectroradiometer (MODIS): evaluationof globalpatterns and comparison with in situ measurements. J. Geophys. Res. 111,G04017.
Zhang, X.Y., Friedl, M.A., Schaaf, C.B., 2009. Sensitivity of vegetation phenologydetection to the temporal resolution of satellite data. Int. J. Remote Sens. 30,2061–2074.
Zhou,Y., Chen, J.,Chen, X.H., et al., 2013. Two importantindicatorswith potential toidentify CaraganaMicrophylla in xilin golgrassland from temporal MODIS data.Ecol. Indic. 34, 520–527.
Zhu, W.Q., Tian, H.Q., Xu, X.F., et al., 2012. Extension of the growing season dueto delayed autumn over mid and high latitudes in North America during1982–2006. Glob. Ecol. Biogeogr. 21, 260–271.
Zhuo,L., Cao, X.,Chen,J., etal., 2007. Assessmentof GrasslandEcological Restoration
Project in Xilin Gol Grassland. Acta Geogr. Sin. 62, 471–480 (in Chinese withabstract in English).