Performance of Random Forest in predicting soil loss based on values calculated by USLE

Soil erosion directly affects agricultural productivity and water resource quality, but estimating soil loss is complex and costly. This study proposes a machine learning (ML) approach to predict soil loss using selected factors from the Universal Soil Loss Equation (USLE) and the Normalized Difference Vegetation Index (NDVI). We applied the Random Forest (RF) algorithm to train and validate two models using different combinations of predictors: (1) NDVI, topographic factor (LS), and land cover/management factor (CP); and (2) NDVI, LS, and soil erodibility factor (K). These variables represent land use, conservation practices, and topographic conditions in the Sorocabuçu River Basin (SRB), part of Brazil’s Atlantic Forest biome with high environmental and socioeconomic value. Soil loss was classified into three classes (in ton/ha): low (0–10.0), moderate (10.1–50.0), and high (≥50.1). A total of 3348 samples were randomly selected and proportionally distributed to reflect class representation across the study area. We used a 70/30 train-test split and standardized parameters (50 trees and four variables per node) to enable reproducibility. The model using NDVI, LS, and CP achieved 93.43% accuracy with a kappa index of 0.90. The performance was especially strong for the low-loss class, the most prevalent in the area. The second model using NDVI, LS, and K achieved 97.14% accuracy with a kappa index of 0.90, showing excellent results, particularly for the high-loss class, which poses the greatest environmental risk. These models prove effective in identifying areas at risk of severe erosion using fewer, more accessible parameters. The approach offers a scalable and practical tool for decision-makers, environmental managers, and public agencies to monitor and mitigate soil degradation, particularly in sensitive and ecologically important regions.