Quality assessment of MODIS vegetation continuous fields of tree cover over France
Résumé
The MODIS VCF is the unique product that enable to map tree cover at globalscale and assess its dynamic for almost two decades. It is used for a wide rangeof environmental studies and provide essential knowledge on the distribution oftrees on earth. The product has already been validated in different areas ofthe world. However, information on its accuracy remains very patchy and theunderstanding of the spatial distribution of errors limited.In this study, we investigate the accuracy of the MODIS VCF percent treecover layer (collection 5, 250-m spatial resolution) over France using the mostdetailed topographical vector database available at the national scale (BDTopo,French mapping agency IGN; minimum mapping unit of 500 m2). We alsocompare the VCF data with a new national land cover map (the OSO product)based on a fully automatic processing 191 chain using Sentinel-2 image time series.In a first time, we estimate for each VCF pixel (n = 5,909,620), the corresponding tree cover percentage of the two reference datasets, while ensuring consistency between the production years. Then, a set of multiple independentvariables are computed including some environmental factors (such as elevationand slope), the level of fragmentation of woody vegetation, and the land covercomposition in the VCF pixel. Finally, uncertainty is quantified from classicalstatistical indicators (Root-Mean-Squared Error, Mean Bias Error, Mean Abso-lute Error) and differences in tree cover are mapped, to understand the spatialpattern and possible sources of errors.Preliminary results show significant differences between the VCF tree coverand the reference products despite the high positive correlation (Spearman’srho correlation = 0.75 for VCF vs BDTopo, and rho = 0.74 for VCF vs OSO;p<0.001). Negative bias with high variability was measured relative to BDTopo(MBE = -12.04 ± 22.41%; RMSE = 25.44%) and OSO (MBE = -6.59 ± 22.34%;RMSE=23.29%). Understimation of VCF tree cover is mainly observed from25% to 75% of cover. Overestimation also appears, mainly in pixels dominatedby grasslands. These findings corroborate previous studies, except for sparselytree areas. No relationship is observed with the quality level of the VCF product.Differences are only slightly explained by slope and elevation.This study contributes to the effort of independant validation of the VCFproduct and may help to better understand the classification errors in variouslandscape contexts.