Modeling Soil Erosion Susceptibility Using Machine Learning Techniques: Rud-e-Faryab Basin, Iran

Abstract

Soil erosion poses a significant threat to water and soil resources, affecting agricultural productivity, infrastructure, and environmental stability. This study models erosion susceptibility in the Rud-e-Faryab basin (Bushehr province, Iran) using the BIOMOD-2 package in R (an ensemble of 10 machine learning algorithms) applied to 10 important environmental variables. Field data on erosion events were used to train and validate the model, and the performance of the model was evaluated using ROC, KAPPA, and TSS coefficients. The results indicate different accuracies for different erosion types, highlighting the GLM, RF, ANN, SRE, and MARS models. Geological formation, slope, and soil resources were found to be the most important factors for erosion susceptibility in the study. Key innovations of this study include (1) the first-time adaptation of the BIOMOD-2 package for soil erosion assessment, (2) the introduction of a stability analysis framework with 10 repeated model runs to test reproducibility, and (3) a comprehensive comparison of 10 machine learning models to identify context-specific optimal approaches. These contributions provide a robust, replicable framework for erosion risk mapping that is particularly valuable in regions with sparse data, and provide actionable insights for sustainable land use planning and resource management.