[ The methods used by the transfer functions to estimate the biophysical variable values ]
 
Warning: the following comments are partial. For more details, please read the processing report.
Site Country Landcover Campaign
date		 AVE* REG**
ρ=reflectance
log=logarithm		 LUT
***		 Comments
Aek Loba Indonesia plam tree plantation 2001/05 - ρ: LAI - LAI map retrieved using the linear NIR-LAI relationship
Les Alpilles 1 & 2 France crops 2001/03 - ρ: LAI - only the multiple regression was applied
2002/07 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG on log(ρ) provides good results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected
Barrax Spain cropland 2003/07 class 2 ρ: LAI,LAI57,
fCover,fAPAR		 - very similar results between REG on ρ and REG on log(ρ) in terms of cross-validation RMSE, but the number of ESUs with weights < 0.7 is higher when using the log(ρ)
Camerons Australia broadleaf forest 2004/03 class 3 ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - the relationship between biophysical variables and NDVI is consistent and REG on ρ provides good results; class 3 corresponds to forest roads (LAI = 0)
Chilbolton England crops ands forest 2006/06 - ρ: LAI,fCover - REG method on ρ provides better results for all the biophysical variables
Concepción Chile mixed forest 2003/01 - ρ: LAI,LAI57,
fAPAR		 - very similar results between REG on ρ and REG on log(ρ), but the number of ESUs with weights < 0.7 is higher when using the log(ρ)
Counami French Guiana tropical forest 2001/09 classes 2,3,4,5 - - no relationship between NDVI and biophysical variables; AVE method is applied (class 1 = clouds)
2002/10 classes 1,3 - - no relationship between NDVI and biophysical variables; the average value of the ESUs is representative since the Counami site is very homogeneous (class 2 = clouds)
Demmin Germany crops 2004/06 all classes - - no relationship between NDVI and biophysical variables; AVE method is applied, but the classification is derived from a land cover map
Donga Benin grassland 2005/06 classes 1,2 - - no satellite high spatial resolution image is available; a land cover map is used to estimate the biophysical variable values, but the results are not consistent; AVE method is applied to classes 1 and 2; spatial aggregation is necessary: medium spatial resolution biophysical variable maps are produced in order to reduce the estimation errors
Fundulea Romania crops 2001/03 - ρ: LAI - only the multiple regression was applied
2001/05 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG on log(ρ) provides slightly better results but it provides unrealistic biophysical variable values; REG on ρ is thus selected
2002/05 class 2 ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG method on ρ provides better results for all the biophysical variables; AVE method is applied to class 2 (peas)
2003/05 class 1 ρ: LAI,LAI57,
fCover,fAPAR		 - similar results between REG on ρ and REG on log(ρ) in terms of cross-validation RMSE
Gilching Germany crops and forest 2002/07 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG method provides better results in terms of cross-validation RMSE for all the variables; close results between REG on ρ and REG on log(ρ)
Gnangara Australia broadleaf forest 2004/03 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - the relationship between biophysical variables and NDVI is not very good, but REG on ρ provides satisfactory results
Gourma Mali grassland 2000/08 - - - the site is relatively homogeneous in terms of LAI and NDVI; the average biophysical variable values are representative over the whole site; no high resolution biophysical variable map
2001/10 - - - the site is relatively homogeneous in terms of LAI; the average biophysical variable values are representative over the whole site; no high resolution biophysical variable map
Haouz Morocco cropland 2003/03 - log(ρ): LAI,LAI57,
fCover,fAPAR		 - the number of ESUs with weights < 0.7 is lower using the log(ρ); LUT method provides systematically higher RMSE value than for REG (weighted RMSE for REG)
Hirsikangas Finland forest 2003/08 class 5 log(ρ): LAI,fCover - LUT method provides systematically higher RMSE value than for REG
2004/07 - ρ: LAI, fCover - the results using the log(ρ) are slightly better, but the estimations are not consistent (very high values on the whole site)
2005/06 - log(ρ): LAI, ρ: fCover - for fCover, the results using the log(ρ) are slightly better, but the estimations are not consistent (very high values on the whole site)
Hombori Mali grassland 2002/08 - - - the site is relatively homogeneous in terms of LAI and NDVI; the average biophysical variable values are representative over the whole site; no high resolution biophysical variable map
Hyytiälä Finland evergreen forest 2008/07 class 4 ρ: LAI, fCover - REG method provides good results; as the relationship between NDVI and biophysical variables of class 4 (forest) is distinguishable from classes 1 and 3, AVE method is applied; class 2 corresponds to water
Järvselja Estonia boreal forest 2000/07 - ρ: LAI, fCover - REG on log(ρ) provides slightly better results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected
2001/06 - ρ: LAI, fCover - for LAI, REG on log(ρ) provides slightly better results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected
2002/06 class 2 ρ: LAI, fCover - for LAI, REG on log(ρ) provides slightly better results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected
2003/07 - log(ρ): LAI;
ρ: fCover		 - REG method provides better results in terms of cross-validation RMSE; for LAI, the results using the log(ρ) are slightly better; a single multiple regression was applied for fCover
2005/06 - ρ: LAI, fCover - for LAI, REG on log(ρ) provides slightly better results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected for LAI and fCover
2007/04 - ρ: LAI, fCover - REG on log(ρ) provides slightly better results but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected for LAI and fCover
2007/07 - ρ: LAI, fCover - REG on ρ provides the best results for LAI and fCover; the class 1 corresponds to clouds
Laprida Argentina grassland 2001/11 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG provides slightly better results (LAItrue, fCover, fAPAR) on log(ρ) but the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; REG on ρ is thus selected (class 2 = cloud and its shade)
2002/10 classes 1,2,3,4 - - no relationship between NDVI and biophysical variables; the average value of the ESUs is representative of the classes 1, 2 and 3; the class 4 has a specific behaviour
Larose Canada boreal forest 2003/08 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - close results between REG on ρ and REG on log(ρ)
Larzac France grassland 2002/07 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - very close results between REG on ρ and REG on log(ρ)
Nezer France pine forest 2000/07 - ρ: LAI,fCover - although the relationship between NDVI and biophysical variables is not very consistent, REG method on ρ provides satisfactory results
2001/04 - ρ: LAIeff,fCover - for fCover, REG on log(ρ) provides slightly better results but it is not selected: the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; therefore, REG on ρ is selected
2001/06 - ρ: LAIeff,fCover - REG on log(ρ) provides slightly lower RMSE values but it is not selected, because the biophysical variable maps obtained from it are not pertinent (very high values); REG on ρ is thus selected
2002/04 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - for LAI variables and fAPAR, REG on log(ρ) provides slightly better results but it is not selected because: for LAI, the tranfer function creates coplanar points which do not allow to determine the ‘strict’ and ‘large’ convex hulls; for fAPAR, it provides negative values; REG on ρ is thus selected
Plan-de-Dieu France crops 2004/07 class 1 ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - the relationship between biophysical variables and NDVI is consistent and REG on ρ provides satisfactory results; class 1 corresponds to woods
Puéchabon France mediterranean forest 2001/06 class 3 ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - the relationship between biophysical variables and NDVI is consistent and REG on ρ is applied; class 3 corresponds to a quarry (LAI = 0)
Romilly-sur-Seine France cropland 2000/06 - - - LAI map retrieved using collocated kriging
Rovaniemi Finland forest 2004/06 - ρ: LAI,fCover - even if the relationship between biophysical variables and NDVI is not consistent, REG on ρ provides satisfactory results
2005/06 class 2 ρ: LAI,fCover - no relationship between NDVI and biophysical variables, but REG on ρ provides good results in terms of cross-validation RMSE
Sonian forest Belgium forest 2004/07 classes 1,2,4,5 - - no relationship between NDVI and biophysical variables; note that the average value of the ESUs is representative of the classes 1, 2 and 5; the class 4 has a specific behaviour and class 3 = cloud
Sud-Ouest France crops 2002/07 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - REG on ρ method provides better results in terms of cross-validation RMSE for all the biophysical variables
Turco Bolivia cropland 2001/07 classes 3,4 - - no relationship between NDVI and biophysical variables; AVE method is applied; class 1 corresponds to water (LAI = 0) and class 2 corresponds to riverbed (not sampled)
2002/08 - ρ: LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - very close results between REG on ρ and REG on log(ρ)
2003/04 - ρ: LAI57,fCover,
fAPAR		 - very close results between REG on ρ and REG on log(ρ)
Wankama Niger grassland 2005/06 classes 1,2,3,4 - - no relationship between NDVI and biophysical variables; AVE method is applied; note that the biophysical variable values are very low
Zhang Bei China pastures 2002/08 - ρ: LAIeff,LAItrue,
LAI57eff,LAI57true,
fCover,fAPAR		 - the relationship between biophysical variables and NDVI is consistent and REG on ρ is applied
*AVE: this method is used if the number of ESUs belonging to the class is too low. The transfer function consists only in attributing the average value of the biophysical variable measured on the class to each pixel of the SPOT image belonging to the class.

**REG: if the number of ESUs is sufficient, multiple robust regression between ESUs reflectance (or Simple Ratio) and the considered biophysical variable can be applied: we used the ‘robustfit’ function from the matlab statistics toolbox. It uses an iteratively re-weighted least squares algorithm, with the weights at each iteration computed by applying the bisquare function to the residuals from the previous iteration. This algorithm provides lower weight to ESUs that do not fit well. The results are less sensitive to outliers in the data as compared with ordinary least squares regression. At the end of the processing, three errors are computed: classical root mean square error (RMSE), weighted RMSE (using the weights attributed to each ESU) and cross-validation RMSE (leave-one-out method).

***LUT: if the number of ESUs is sufficient, the look-up table is built using ESUs reflectances and the corresponding measured biophysical variable. For a given pixel, a cost function is computed as the sum of the square difference between the pixel reflectances and the ESU reflectances over the 3 or 4 bands, divided by the standard deviation computed on ESU reflectances. The result of the cost function is sorted in ascending order, and the biophysical variable estimated for the given pixel is computed as the mean value of the first n ESUs providing the lowest value of the cost function. Different values of n are considered to get the lowest cost function. This method is reliable only if the ESU NDVI distribution is quite comparable with the whole site NDVI distribution.