Optical–SAR Data Fusion and Machine Learning for Soil Erosion Susceptibility Mapping in the Rukuru Basin, Malawi

Abstract Soil erosion threatens agricultural productivity and sediment-related hazards in Sub-Saharan Africa. This study integrates optical and synthetic aperture radar (SAR) remote sensing with Random Forest machine learning to map soil erosion susceptibility in the South and North Rukuru River Basin, Malawi. Predictor variables, including slope, topographic wetness index, drainage density, NDVI, and SAR VV backscatter, were fused to capture spatial heterogeneity. The integrated model outperformed single-sensor configurations, achieving RMSE = 0.124, MAE = 0.091, R² = 0.872, AUC = 0.97 (95% CI: 0.955–0.983), and confusion matrix metrics of accuracy = 0.914, sensitivity = 0.923, specificity = 0.905, precision = 0.906, and F1-score = 0.914. Optical-only and SAR-only models yielded lower AUCs (0.84 and 0.88), emphasizing the benefit of multi-sensor integration. Results show that 21.4% of the basin is high to very high susceptibility, concentrated in northern uplands with steep slopes (> 15°), while 43.9% falls within low to very low susceptibility, largely in floodplains and vegetated valley bottoms. Land cover change (2000–2025) reveals that 68% of woodland-to-cropland and 61% of shrubland-to-cropland conversions occurred in moderate to high susceptibility zones, increasing mean erosion probability by + 0.32 and + 0.27. The study demonstrates that multi-sensor Random Forest models provide robust, spatially explicit erosion susceptibility predictions, offering essential guidance for targeted soil conservation and sustainable land management in rapidly transforming tropical catchments.