Large underestimation of smallholder agriculture in the tropics by classic remote sensing products: evidenced from the Kasaï River Basin

Abstract The Kasaï River Basin (KRB), the largest tributary system of the Congo basin, is undergoing rapid land-use transformation driven by population growth and expanding economic activities. However, existing global and regional land-cover products do not fully capture this highly heterogeneous, smallholder-dominated landscape or its temporal dynamics. Here, we develop a regionally calibrated 30 m resolution land-cover dataset for the KRB for 2017–2024 by combining 3,734 reference observations across five classes (urban, water, forest, savanna/grassland, cropland) with a random forest classifier trained on multi-source Earth observation datasets encoded as embeddings by Google’s AlphaEarth Foundations. The resulting classification achieves 96%–100% accuracy for urban and water classes and 74%–87% for forest, savanna/grassland, and cropland classes, substantially outperforming current global products (30%–65%). This improvement results from using a large region-specific training dataset and a modeling strategy designed to capture land surface phenology; this enhances the detection of inter-annual dynamics in rotational and intermittent smallholder agriculture, characteristic for many tropical regions such as the KRB. We estimate that in 2024 cropland covered 16.2% of the KRB, while forest and savanna/grassland occupied 41.1% and 41.5%, respectively—values that contrast sharply with those from existing products, which report cropland extents of only 0.1%–4.7%. Transition analysis further reveals that over 53 128 km 2 of forest and grassland in the KRB have been converted to cropland between 2017 and 2024. These results highlight the need for regionally calibrated land-cover products in heterogeneous tropical landscapes. By resolving the systematic biases present in existing global products, the dataset produced here provides an accurate, high-resolution benchmark that underpins environmental modeling, food security analysis, and land-use decision-making in the KRB and other tropical regions.