temporally-dense, high-resolution lai estimation using...

1
Temporally-dense, high-resolution lai estimation using multi-sensor fusion Josh Gray and Conghe Song University of North Carolina at Chapel Hill Project at a Glance Objective: Create a 30 m resolution map of LAI at a 2-week interval for a study-site in the NC piedmont. Method: A novel multi-sensor fusion approach blending Ikonos, Landsat and MODIS imagery. Texture is a surrogate for canopy structure and is a complimen- tary estimator to spectral indices. Multiple regression using multiple SVI and texture measures can leverage complimentary information. AIC based model-selection can reveal alternative, good models. Our model is simple to apply and successful in producing rea- sonable maps of LAI at an arbitrary temporal resolution. Motivation Ecosystem process simulations on local- to regional-scales require high spatial and temporal resolution maps of LAI to achieve reliable simulation results. The tradeoff between achievable spatial and temporal resolutions from remote sensing platforms has stimulated research into multi-sensor fusion models. We propose a novel method of LAI estimation which com- bines complimentary information from remotely sensed imagery spanning three orders of magnitude in spatial resolution. Texture information from high-resolution Ikonos imagery and spectral information from moderate resolution Landsat TM imagery is used to capture the fine-scale spatial variability in LAI, and coarser resolution MODIS data is used to extract the temporal signal of phenological change. This study demonstrates that the use of multiple SVI as well as texture information can significantly improve LAI estimation when compared with conventional approaches. We used the model to construct 30 m spatial resolution LAI maps at a 2-week interval for a study-area in the NC piedmont. The model is simple to run, not data intensive, and produces reasonable estimates of the time evolution of LAI throughout a growing season. 0 750 Meters 0 2 Kilometers LAI Plot Ameriflux Site Figure 1: Left Panel: May, 19 2009 Landsat TM composite of study-area (10 km x 10 km ex- tent of Ikonos image). Right Panel: The Duke Forest research area where ground-observations of LAI were made. Sampling plots are shown on the panchromatic Ikonos image. Model Overview Separate empirical models for conifer and deciduous forest types are fit to ground observations using a combination of spectral and spatial infor- mation from Landsat and Ikonos imagery and used to generate a map of maximum LAI (Eq. 3). A deciduousness parameter is calculated (Eq. 6) and used to mix LAI contributions from conifer and deciduous vegetation within a given pixel (Eq. 1). A phenological function (Eq. 4 and 5) is used to adjust an individual pixel’s LAI between it’s minimum and maximum values for an arbitrary day, t, during the growing season (Eq. 2). LAI(t) = (1 - Ω)LAI c (t) + ΩLAI d (t) (1) LAI c,d (t) = LAI min c,d + f c,d (t) LAI max c,d - LAI min c,d (2) LAI max c,d = β 0 + β 1 SVI 1 + β 2 SVI 2 + β 3 TEX (3) f c,d (t) = MODSVI c,d (t) [0, 1] (4) MODSVI c,d (t)= 1 1+ e a-bt - 1 1+ e a -b t g + h (5) Ω= NDVI summer - NDVI winter NDVI summer + NDVI winter [0, 1] (6) Empirical LAI Model We use two innovative techniques to fit an empirical model to ground measured LAI: 1) we utilize image texture from high-resolution imagery to compliment multiple SVIs in a multiple regression framework, and 2) we adopt the information theoretic criteria approach to model selection. Ground-based estimates of effective LAI were obtained for a mixture of homogenous evergreen and deciduous plots (n = 33) using indirect optical techniques (LAI-2000). Field campaigns were conducted in late June and July of 2009. Landsat TM imagery was obtained to coincide with the field collections and atmospherically corrected using a dark-object subtraction approach with downwelling diffuse radiation calculated using 6S. 1 m 0 5 10 15 20 25 0e+00 2e+04 4e+04 6e+04 8e+04 1e+05 h (m) a ^ (h) 1 m 2 m 3 m 5 m 1 m 0 2 4 6 8 10 12 0 10000 20000 30000 40000 50000 h (m) a ^ (h) 1 m 2 m 3 m 5 m 3 m 3 m 5 m 5 m Figure 2: Semivariograms for deciduous (left) and pine (right) calculated at four spatial resolu- tions on 250x250 pixel samples from the Ikonos image. The effect of regularization is to lower the overall variance in the sample (sill) and increase the range. Note the significantly smaller spatial scale of the pine spatial pattern. Eight multiple regression models were proposed to fit the observed LAI data. We allowed models to have two SVI predictors, the first being either the simple ratio (SR), or reduced simple ratio (RSR), and the second being either the structural index (SI), or enhanced vegetation index (EVI). Ad- ditionally, half of the models included a texture predictor (VAR). We chose to use a first-order measure of image texture, windowed variance, although alternatives such as GLCM features have also demonstrated effectiveness in estimating canopy structure. We determined the window size based on empirical semivariograms at various pixel resampling sizes (Fig. 2) as well as images of windowed variance over the study area (Fig. 3). The models were fit and ranked via AIC (Table 1). Texture information leads to a significant improvement in estimating deciduous LAI, whereas conifer LAI is best predicted using SVI alone (Fig. 4). Windowed Variance 5e5 0 Windowed Variance 4e5 0 Windowed Variance 3e5 0 Figure 3: Variance calculated on 1 m spatial resolution Ikonos panchromatic image at three window size: 13, 21 and 33 pixels. Note the differences in color-scale for each. Table 1: Candidate models ranked by AICc and associated parameter estimates. All models have effective LAI (L e ) as the dependent variable. Sig: ***: <0.001, **: <0.01, *: <0.05, ·: <0.1 Model Parameters Rank AICc AIC Weight R 2 SEV SR c SR d EVI c EVI d VAR c VAR d 1 87.4 0.71 0.73 1.0e1 *** 2.9 · -1.0e2 *** -4.8e1 * -1.6e-5 1.1e-4 ** REV RSR c RSR d EVI c EVI d VAR c VAR d 2 90.0 0.20 0.71 5.2 *** 1.7 -3.0e1 * -1.8e1 * -3.2e-5 9.6e-5 ** SE SR c SR d EVI c EVI d 3 92.5 0.05 0.61 1.0e1 *** 4.3e-1 -1.0e2 *** -8.3 RE RSR c RSR d EVI c EVI d 4 93.7 0.03 0.59 5.1 *** 2.0e-1 -3.4e1 * -3.7 RS RSR c RSR d SI c SI d 5 100.4 <0.01 0.50 1.3 4.8e-1 3.7 0.7 RIV RSR c RSR d SI c SI d VAR c VAR d 6 101.1 <0.01 0.59 3.3 -1.8e-1 8.3e-1 -2.2 -5.5e-5 6.0e-5 SIV SR c SR d SI c SI d VAR c VAR d 7 101.5 <0.01 0.59 -1.2e-1 -4.6e-1 7.4 ** 2.4e-1 -3.6e-5 6.8e-5 · SI SR c SR d SI c SI d 8 189.8 <0.01 0.33 1.7 *** 2.0e-1 · 8.3e-1 ** -6.2e-2 2 3 4 5 6 2 3 4 5 6 Predicted L e A Evergreen Deciduous 2 3 4 5 6 2 3 4 5 6 Predicted L e B Evergreen Deciduous Figure 4: Comparison of observed and predicted L e using models SE (A) and SEV (B). Model SEV contains a texture estimator which leads to a significant improvement in model fit for deciduous stands. Phenology We fit the difference logistic function (Eq. 5) to MODIS NDVI. The VI Reliability Index was used to limit calculations to reliable pixels only, and the MOD12 Landcover product (Type 1) was used to construct separate time-series of deciduous and evergreen vegetation. Daily mean values of NDVI were calculated by using the Composite DOY product. Further filtering was necessary to remove outliers and dates for which there were less than 20 pixels available for calculation of the mean (Fig. 5). The re- sulting time-series were fit using weighted nonlinear least squares methods where the weights are determined as the reciprocal of the sample mean (i.e., N i 2 ). It was found that the evergreen and deciduous time-series had identical temporal trajectories, so we use the deciduous model for both evergreen and deciduous LAI estimation. 0.0 0.2 0.4 0.6 0.8 2009ï02ï14 2009ï04ï28 2009ï07ï10 2009ï09ï21 2009ï12ï03 2010ï02ï14 MOD12 1 Decid NDVI Used for fit Omited: N<20 Omited: Outlier 0.4 0.5 0.6 0.7 0.8 0.9 2009ï02ï14 2009ï04ï28 2009ï07ï10 2009ï09ï21 2009ï12ï03 2010ï02ï14 MOD12 1 Decid NDVI Fitting Weight Percentile <25% 25%ï75% >75% Fitted Curve Figure 5: Left panel: Mean NDVI for MOD12Q1 Type 1 Deciduous forests indicating points omitted from fitting procedure by outlier detection (Tukey IQR method) or too few pixels used in calculation. Right panel: Full range of fitted data. Point sizes indicate the relative weight used in the nonlinear least squares fit. Dashed line is fit. Results We used the full model to estimate maps of LAI at a 2-week interval for the study-area. Fig. 6 shows four such maps at select temporal intervals. Two Ameriflux sites within the study-area have a daily record of LAI for deciduous and conifer plots. We compared these records to our daily es- timates of LAI at these plots (Fig. 7). Our model does a good job of capturing both the temporal trend as well as the magnitude of LAI at these plots. However, for the deciduous stand, we note a longer senescence period for our model than is indicated by the ground observations. This is likely due to understory herbaceuous and evergreen vegetation which becomes visible to the satellite sensor as the overstory canopy sheds its leaves. LAI 12 0 Feb. 14 April 26 July 1 Nov. 1 Figure 6: LAI map of the study area at four select dates throughout the growing season. Mini- mum, green-up, peak and senescence phases of the vegetation are depicted. 0 1 2 3 4 5 6 7 Decid LAI Mar May Jul Sep Nov Jan AmeriFlux Data Modeled 0 1 2 3 4 5 6 7 Pine LAI Mar May Jul Sep Nov Jan AmeriFlux Data Modeled Figure 7: Validation data from Duke Ameriflux towers. Conclusions and Future Directions We demonstrate that texture indices lead to significant improvements in the estimation of LAI, particularly for deciduous vegetation. Texture acts as a surrogate for canopy structure and contains information about crown size and stocking density. We show that empirical variograms on representative textures are useful for determining texture parameters such as window size and resampling resolution. The successive filtering steps we applied to MODIS NDVI 16-day com- posites allow for the extraction of smooth time-series suitable for de- termining stable parameters of phenological functions. That evergreen and deciduous time-series were temporally identical indicates that there is a substantial amount of mixed or misclassified pixels in the MOD12 landcover product, and that it may be unsuitable for this application. Overall, our method represents a very simple method for fusing compli- mentary remote sensing information from three distinct spatial resolu- tions. This method has demonstrated its effectiveness in representing the magnitude and temporal signature of LAI production. Future improvements should focus on improving the phenological time- series data by using improved landcover classifications. Additional val- idation, particularly in mixed stands, will improve confidence in our modeling approach.

Upload: others

Post on 21-Aug-2020

11 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Temporally-dense, high-resolution lai estimation using ...lcluc.umd.edu/sites/default/files/lcluc_documents/Gray_poster_0.pdf · collections and atmospherically corrected using a

Temporally-dense, high-resolution lai estimation using multi-sensor fusionJosh Gray and Conghe Song

University of North Carolina at Chapel Hill

Project at a Glance•Objective: Create a 30 m resolution map of LAI at a 2-weekinterval for a study-site in the NC piedmont.

•Method: A novel multi-sensor fusion approach blendingIkonos, Landsat and MODIS imagery.

• Texture is a surrogate for canopy structure and is a complimen-tary estimator to spectral indices.

•Multiple regression using multiple SVI and texture measures canleverage complimentary information.

•AIC based model-selection can reveal alternative, good models.

•Our model is simple to apply and successful in producing rea-sonable maps of LAI at an arbitrary temporal resolution.

Motivation

Ecosystem process simulations on local- to regional-scales require highspatial and temporal resolution maps of LAI to achieve reliable simulationresults. The tradeo! between achievable spatial and temporal resolutionsfrom remote sensing platforms has stimulated research into multi-sensorfusion models. We propose a novel method of LAI estimation which com-bines complimentary information from remotely sensed imagery spanningthree orders of magnitude in spatial resolution. Texture information fromhigh-resolution Ikonos imagery and spectral information from moderateresolution Landsat TM imagery is used to capture the fine-scale spatialvariability in LAI, and coarser resolution MODIS data is used to extractthe temporal signal of phenological change. This study demonstrates thatthe use of multiple SVI as well as texture information can significantlyimprove LAI estimation when compared with conventional approaches.We used the model to construct 30 m spatial resolution LAI maps at a2-week interval for a study-area in the NC piedmont. The model is simpleto run, not data intensive, and produces reasonable estimates of the timeevolution of LAI throughout a growing season.

0 750 Meters0 2 Kilometers

LAI PlotAmeriflux Site

Figure 1: Left Panel: May, 19 2009 Landsat TM composite of study-area (10 km x 10 km ex-tent of Ikonos image). Right Panel: The Duke Forest research area where ground-observationsof LAI were made. Sampling plots are shown on the panchromatic Ikonos image.

Model Overview

Separate empirical models for conifer and deciduous forest types are fitto ground observations using a combination of spectral and spatial infor-mation from Landsat and Ikonos imagery and used to generate a map ofmaximum LAI (Eq. 3). A deciduousness parameter is calculated (Eq. 6)and used to mix LAI contributions from conifer and deciduous vegetationwithin a given pixel (Eq. 1). A phenological function (Eq. 4 and 5) is usedto adjust an individual pixel’s LAI between it’s minimum and maximumvalues for an arbitrary day, t, during the growing season (Eq. 2).

LAI(t) = (1! ")LAIc(t) + "LAId(t) (1)

LAIc,d(t) = LAIminc,d + fc,d(t)

!LAImax

c,d ! LAIminc,d

"(2)

LAImaxc,d = !0 + !1SVI1 + !2SVI2 + !3TEX (3)

fc,d(t) = MODSVIc,d(t) " [0, 1] (4)

MODSVIc,d(t) =

#1

1 + ea!bt! 1

1 + ea#!b#t

$g + h (5)

" =NDVIsummer ! NDVIwinterNDVIsummer + NDVIwinter

" [0, 1] (6)

Empirical LAI Model

We use two innovative techniques to fit an empirical model to groundmeasured LAI: 1) we utilize image texture from high-resolution imageryto compliment multiple SVIs in a multiple regression framework, and 2)we adopt the information theoretic criteria approach to model selection.Ground-based estimates of e!ective LAI were obtained for a mixture ofhomogenous evergreen and deciduous plots (n = 33) using indirect opticaltechniques (LAI-2000). Field campaigns were conducted in late June andJuly of 2009. Landsat TM imagery was obtained to coincide with the fieldcollections and atmospherically corrected using a dark-object subtractionapproach with downwelling di!use radiation calculated using 6S.

1 m

0 5 10 15 20 25

0e+0

02e

+04

4e+0

46e

+04

8e+0

41e

+05

h (m)

^ (h)

1 m 2 m 3 m 5 m 1 m

0 2 4 6 8 10 12

010

000

2000

030

000

4000

050

000

h (m)

^ (h)

1 m 2 m 3 m 5 m

3 m 3 m

5 m 5 m

Figure 2: Semivariograms for deciduous (left) and pine (right) calculated at four spatial resolu-tions on 250x250 pixel samples from the Ikonos image. The e!ect of regularization is to lower theoverall variance in the sample (sill) and increase the range. Note the significantly smaller spatialscale of the pine spatial pattern.

Eight multiple regression models were proposed to fit the observed LAIdata. We allowed models to have two SVI predictors, the first being eitherthe simple ratio (SR), or reduced simple ratio (RSR), and the second beingeither the structural index (SI), or enhanced vegetation index (EVI). Ad-ditionally, half of the models included a texture predictor (VAR). We choseto use a first-order measure of image texture, windowed variance, althoughalternatives such as GLCM features have also demonstrated e!ectivenessin estimating canopy structure. We determined the window size based onempirical semivariograms at various pixel resampling sizes (Fig. 2) as wellas images of windowed variance over the study area (Fig. 3). The modelswere fit and ranked via AIC (Table 1). Texture information leads to asignificant improvement in estimating deciduous LAI, whereas conifer LAIis best predicted using SVI alone (Fig. 4).

WindowedVariance

5e5

0

WindowedVariance

4e5

0

WindowedVariance

3e5

0

Figure 3: Variance calculated on 1 m spatial resolution Ikonos panchromatic image at threewindow size: 13, 21 and 33 pixels. Note the di!erences in color-scale for each.

Table 1: Candidate models ranked by AICc and associated parameter estimates. All models havee!ective LAI (Le) as the dependent variable. Sig: ***: <0.001, **: <0.01, *: <0.05, ·: <0.1

Model Parameters Rank AICc AIC Weight R2

SEV SRc SRd EVIc EVId VARc VARd 1 87.4 0.71 0.731.0e1$$$ 2.9· -1.0e2$$$ -4.8e1$ -1.6e-5 1.1e-4$$

REV RSRc RSRd EVIc EVId VARc VARd 2 90.0 0.20 0.715.2$$$ 1.7 -3.0e1$ -1.8e1$ -3.2e-5 9.6e-5$$

SE SRc SRd EVIc EVId 3 92.5 0.05 0.611.0e1$$$ 4.3e-1 -1.0e2$$$ -8.3

RE RSRc RSRd EVIc EVId 4 93.7 0.03 0.595.1$$$ 2.0e-1 -3.4e1$ -3.7

RS RSRc RSRd SIc SId 5 100.4 <0.01 0.501.3 4.8e-1 3.7 0.7

RIV RSRc RSRd SIc SId VARc VARd 6 101.1 <0.01 0.593.3 -1.8e-1 8.3e-1 -2.2 -5.5e-5 6.0e-5

SIV SRc SRd SIc SId VARc VARd 7 101.5 <0.01 0.59-1.2e-1 -4.6e-1 7.4$$ 2.4e-1 -3.6e-5 6.8e-5·

SI SRc SRd SIc SId 8 189.8 <0.01 0.331.7$$$ 2.0e-1· 8.3e-1$$ -6.2e-2

2 3 4 5 6

23

45

6

Pred

icted

Le

A

EvergreenDeciduous

2 3 4 5 6

23

45

6

Pred

icted

Le

B

EvergreenDeciduous

Figure 4: Comparison of observed and predicted Le using models SE (A) and SEV (B). ModelSEV contains a texture estimator which leads to a significant improvement in model fit for deciduousstands.

Phenology

We fit the di!erence logistic function (Eq. 5) to MODIS NDVI. The VIReliability Index was used to limit calculations to reliable pixels only, andthe MOD12 Landcover product (Type 1) was used to construct separatetime-series of deciduous and evergreen vegetation. Daily mean values ofNDVI were calculated by using the Composite DOY product. Furtherfiltering was necessary to remove outliers and dates for which there wereless than 20 pixels available for calculation of the mean (Fig. 5). The re-sulting time-series were fit using weighted nonlinear least squares methodswhere the weights are determined as the reciprocal of the sample mean(i.e., Ni/"

2). It was found that the evergreen and deciduous time-serieshad identical temporal trajectories, so we use the deciduous model for bothevergreen and deciduous LAI estimation.

0.0

0.2

0.4

0.6

0.8

2009 02 14 2009 04 28 2009 07 10 2009 09 21 2009 12 03 2010 02 14

MO

D12 1

Dec

id N

DVI

Used for fitOmited: N<20Omited: Outlier

0.4

0.5

0.6

0.7

0.8

0.9

2009 02 14 2009 04 28 2009 07 10 2009 09 21 2009 12 03 2010 02 14

MO

D12 1

Dec

id N

DVI

Fitting Weight Percentile<25%25% 75%>75%Fitted Curve

Figure 5: Left panel: Mean NDVI for MOD12Q1 Type 1 Deciduous forests indicating pointsomitted from fitting procedure by outlier detection (Tukey IQR method) or too few pixels used incalculation. Right panel: Full range of fitted data. Point sizes indicate the relative weight used inthe nonlinear least squares fit. Dashed line is fit.

Results

We used the full model to estimate maps of LAI at a 2-week interval forthe study-area. Fig. 6 shows four such maps at select temporal intervals.Two Ameriflux sites within the study-area have a daily record of LAI for

deciduous and conifer plots. We compared these records to our daily es-timates of LAI at these plots (Fig. 7). Our model does a good job ofcapturing both the temporal trend as well as the magnitude of LAI atthese plots. However, for the deciduous stand, we note a longer senescenceperiod for our model than is indicated by the ground observations. Thisis likely due to understory herbaceuous and evergreen vegetation whichbecomes visible to the satellite sensor as the overstory canopy sheds itsleaves.

LAI

12

0

Feb. 14 April 26

July 1 Nov. 1

Figure 6: LAI map of the study area at four select dates throughout the growing season. Mini-mum, green-up, peak and senescence phases of the vegetation are depicted.

01

23

45

67

Decid

LAI

Mar May Jul Sep Nov Jan

AmeriFlux DataModeled

01

23

45

67

Pine

LAI

Mar May Jul Sep Nov Jan

AmeriFlux DataModeled

Figure 7: Validation data from Duke Ameriflux towers.

Conclusions and Future Directions

•We demonstrate that texture indices lead to significant improvementsin the estimation of LAI, particularly for deciduous vegetation. Textureacts as a surrogate for canopy structure and contains information aboutcrown size and stocking density. We show that empirical variograms onrepresentative textures are useful for determining texture parameterssuch as window size and resampling resolution.

• The successive filtering steps we applied to MODIS NDVI 16-day com-posites allow for the extraction of smooth time-series suitable for de-termining stable parameters of phenological functions. That evergreenand deciduous time-series were temporally identical indicates that thereis a substantial amount of mixed or misclassified pixels in the MOD12landcover product, and that it may be unsuitable for this application.

•Overall, our method represents a very simple method for fusing compli-mentary remote sensing information from three distinct spatial resolu-tions. This method has demonstrated its e!ectiveness in representingthe magnitude and temporal signature of LAI production.

• Future improvements should focus on improving the phenological time-series data by using improved landcover classifications. Additional val-idation, particularly in mixed stands, will improve confidence in ourmodeling approach.