Abstract: Accurate information related to soil texture is essential for understanding key biological, chemical, and hydrological processes. However, soil data is scarce and unevenly distributed, especially in tropical regions, and global soil information products lack regional assessments, leading to high uncertainty. Here, we leveraged an unprecedented dataset of soil observations in a particularly poorly documented region, French Guiana, to identify the drivers of soil texture in such territory, develop digital soil maps of textural components, and provide an independent assessment of existing soil products at a regional scale. Specifically, we employed the random forest (RF) model to predict sand, clay and silt contents from multiple environmental variables describing geology, climate, and geomorphology. Results were evaluated through k-fold random and spatial cross-validation. We used our model to derive a soil texture map for French Guiana. We tested our map and global soil texture maps from the Harmonized World Soil Database (HWSD) and SoilGrids against 72 independent soil profiles collected in the region. Geomorphology and geology were the most important predictors of sand, clay, and silt contents in our model, yielding relatively good predictions (random cross-validation: R2 = 0.54, 0.76, and 0.10; spatial cross-validation: R2=0.26, 0.64 and 0.05; RMSE=10.92%, 6.38%, and 6.50%, for sand, clay, and silt contents, respectively). When evaluated on the independent dataset, both SoilGrids and HWSD exhibited poor performance, characterised by lower R2 (<0.07) and higher RMSE values (> 13%). Furthermore, HWSD and SoilGrids failed to capture the spatial heterogeneity of soil texture in the region, calling for caution when using such global products at local scale. Overall, our study emphasises the need for sustained effort in assembling distributed soil information, as well as meaningful soil predictors at local and regional scales.
Processing Request
No Comments.