Abstract

The Moderate Resolution Imaging Spectroradiometer Vegetation Continuous Fields (MODIS VCF) Earth observation product is widely used to estimate forest cover changes and to parameterize vegetation and Earth system models and as a reference for validation or calibration where field data are limited. However, although limited independent validations of MODIS VCF have shown that MODIS VCF's accuracy decreases when estimating tree cover in sparsely vegetated areas such as tropical savannas, no study has yet assessed the impact this may have on the VCF-based tree cover data used by many in their research. Using tropical forest and savanna inventory data collected by the Tropical Biomes in Transition (TROBIT) project, we produce a series of calibration scenarios that take into account (i) the spatial disparity between the in situ plot size and the MODIS VCF pixel and (ii) the trees' spatial distribution within in situ plots. To identify if a disparity also exists in products trained using VCF, we used a similar approach to evaluate the finer-scale Landsat Tree Canopy Cover (TCC) product. For MODIS VCF, we then applied our calibrations to areas identified as forest or savanna in the International Geosphere-Biosphere Programme (IGBP) land cover mapping product. All IGBP classes identified as "savanna" show substantial increases in cover after calibration, indicating that the most recent version of MODIS VCF consistently underestimates woody cover in tropical savannas. We also found that these biases are propagated in the finer-scale Landsat TCC. Our scenarios suggest that MODIS VCF accuracy can vary substantially, with tree cover underestimation ranging from 0 % to 29 %. Models that use MODIS VCF as their benchmark could therefore be underestimating the carbon uptake in forest-savanna areas and misrepresenting forest-savanna dynamics. Because of the limited in situ plot number, our results are designed to be used as an indicator of where the product is potentially more or less reliable. Until more in situ data are available to produce more accurate calibrations, we recommend caution when using uncalibrated MODIS VCF data in tropical savannas.