Using MODIS and POLDER data to develop a generalized approach for correction of the BRDF effect Eric F. Vermote, Christopher O. Justice Dept of Geography, UMCP Francois-Marie Breon Laboratoire des Sciences du Climat et de lEnvironnement, Unit Mixte de Recherche CEA-CNRS-UVSQ ( This research is part of the NASA supported Land LTDR Project) Introduction The effect of surface anisotropy on remotely sensed satellite data has been the subject of intensive research over the past 20 years The surface

reflectance is described by the Bidirectional Reflectance Distribution Function (BRDF), which is a function of the sun zenith angle s, the view zenith angle v, and both azimuths s and v with respect to a reference direction. In practice, for most applications, the azimuth variations only depend on the relative azimuth=s-v. Directional reflectance observed for an evergreen needle leaf forest from Schaaf et al. The Polarization and Directionality of the Earths Reflectances results from POLDER

Using multi-directional Parasol POLDER data at coarse resolution (6 km) over a large set of representative targets, POLDER showed that simple models with only 3 free parameters permit an accurate representation of the BRDFs. The best results (low RMS residuals) were obtained with the linear Ross-Li-HS model, a version of the Ross-Li model that accounts for the Hot-Spot process s, v , k0 k1 F1 s, v , k 2 F2 s, v , k1 k k0 1 F1 s, v , 2 F2 s, v , k0 k 0

The ability of simple, linear, models to reproduce the BRDF of natural targets opens the way for the correction of directional effects on reflectance time series data (MODIS and AVHRR). However, the question remains as to the choice of the BRDF model, i.e. the determination of its free parameters. The POLDER results toward a generic BRDF Measurements from the Polarization and Directionality of the Earths Reflectances (POLDER) BRDF database, have shown that it is possible to assume a typical BRDF signature on a biome basis and therefore apply apriori correction of the BRDF effect. This approach has been applied successfully on wide-swath data from polar orbiting satellite systems (e.g. AVHRR)

Application to MODIS Surface Reflectance CMG daily data Time series (2000 to 2004) MODIS CMG daily Red and Nir reflectance data over a southern Africa Tropical Savanna site yi 2 yi 2 y day day y in

i 1 i 1 i i 1 dayi 2 dayi Noise( y ) N 2

2 Measure of Perturbation associated with the BRDF A new approach to invert BRDF on times series Classic approach assumes the reflectance does not vary within the inversion time interval and BRDF correction minimizes the classic merit function s, v , k0 k1 F1 s, v , k 2 F2 s, v , 2 i i N i 1k 0 k1 F1 k 2 F2 i

k1 k k0 1 F1 s, v , 2 F2 s, v , k0 k 0 Our new approach allows the reflectance to vary slowly within the interval and minimization of a more complicated merit function t i t i1

1 V F1i R F2i 1 V F1i1 R F2i1 1 VF M N 1 i1 i1 i 1

i 2 i1 1 RF i 1 VF day i1 day i 1 i1 2 RF 2

The equation to be solved is still linear N 1 i i F1 F1 i1 N 1 i i F1 F2 i1 N1

N 1 i i F1 F2 V F1 i1 i1 N 1 N1 R i i i i

F F F 2 2 2 i1

i1 i i with i d day i1 day i 1 i i1 i / i d i F1 i1F1i i F1i1 / i d i F2 i1F2i i F2i1 / i d Time series of normalized reflectance using the classical approach (tropical savanna)

1 V F1 45,0,0 RF2 45,0,0 N 45,0,0 s , v , 1 V F1 s , v , RF2 s , v , Uncorrected Reflectance Data Time series of normalized reflectance using the classical approach N 45,0,0 s, v , 1 V F145,0,0 R F2 45,0,0 1 V F1 s, v , R F2 s, v ,

Time-series of normalized reflectance using the new approach Further improvements allow the V (volume parameter) and R (roughness parameter) to vary as a function of NDVI Red: band 2 Blue: band 1 Improving Correction by Stratifying by Vegetation Amount over Time Results of final BRDF Correction Original NDVI

NDVI computed from classical BRDF approach NDVI computed from new BRDF inversion (V and R fixed) NDVI computed from new BRDF inversion (V and R varies linearly with NDVI) Results for various land covers Savanna Channel 1 Channel 2

NDVI Raw data noise Cor. Data noise Raw data noise Cor. Data noise Raw data noise Cor. Data noise 0.019 (18.8%) 0.004 ( 3.4%) 0.040 (16.4%) 0.005 ( 2.4%) 0.019 ( 4.6%) 0.008 ( 2.3%)

Evergreen forest 0.006 (33.6%) 0.002 (13.0%) 0.063 (20.6%) 0.007 ( 2.5%) 0.016 ( 1.8%) 0.012 ( 1.4%) Deciduous forest 0.011 (23.0%) 0.004 (10.0%) 0.043 (19.7%) 0.010 ( 4.5%) 0.017 ( 3.0%)

0.013 ( 2.3%) Broadleaf crops 0.011 (12.6%) 0.005 ( 6.2%) 0.024 ( 9.4%) 0.011 ( 4.5%) 0.027 ( 5.6%) 0.012 ( 2.8%) Broadleaf crops 0.016 (16.3%) 0.006 ( 7.3%) 0.043 (16.5%)

0.011 ( 4.6%) 0.026 ( 5.7%) 0.023 ( 5.3%) Global NDVI (without BRDF correction) 0.0 0.04 Noise on the NDVI computed using the directional reflectance from MODIS band 1 and 2. Global NDVI (with new BRDF correction) 0.0

0.04 Noise on the NDVI computed using the reflectance corrected for BRDF effect from MODIS band 1 and 2 Global reduction in NDVI noise 0.0 50% NDVI Noise reduction in %. Global map of R and V parameters at the peak NDVI 0.0

0.9 NDVI at the peak V parameter at the peak NDVI 0.0 2.5 R parameter at the peak NDVI -0.05 0.25

Sahara Desert Detail Ahaggar Mtns Tibesti Mtns Air Mountains Surface Reflectance (RGB) Details over Sahara (Roughness) The R parameter is related to aerodynamic surface roughness length (Marticonera et al. POLDER data) 0.25 y = 0.22098 + 0.062443log(Z0) R= 0.80215

We used the dataset of roughness length collected by Greeley et al. for Namibia, Death Valley and Lunar Lake U.S.A. and the dataset collected by Marticorena et al. for an arid surface in southern Tunisia. R parameter [unitless] 0.2 0.15 Excluding sites with substantial vegetation cover, we compared the R parameter derived from this study to the aerodynamic roughness length Z0. The relationship derived is close to

the one derived by Marticorena et al. i.e. (0.277+0.052log(Z0)) 0.1 0.05 0 0.0001 0.001 0.01 Roughness length Z0 [cm]

0.1 1 Details over Europe (Roughness) -0.05 0.25 R parameter % Tree cover Hansen et al. (2002) 0

80% A Tale of Two Cities London/Justice Paris/Vermote -0.05 0.25 R parameter High Roughness Associated with Major Cities Conclusions A new approach has been developed and tested to

correct daily time-series of reflectance data for the BRDF effect (using a database of coefficient V and R that only depend on NDVI) (paper in preparation) The NDVI after the new BDRF correction is greatly improved (factor 2 reduction) for a large percentage of the land cover types as compared to non-corrected data Once the database (effectively a time varying map of R an V) is developed - the correction could be applied to other similar time-series data sets without deriving the BRDF The V and R coefficients themselves also could be used in other applications e.g. R could be used for Aerodynamic roughness and land cover characterization We intend to use the approach in the LTDR project to correct AVHRR and MODIS Surface Reflectance time series for the BRDF effect.