A New Approach for Smart Soil Erosion Modeling: Integration of Empirical and Machine-Learning Models

Despite advances in artificial intelligence modelling, the lack of soil erosion data and other watershed information is still one of the important factors limiting soil-erosion modelling. Additionally, the limited number of parameters and the lack of evaluation criteria are major disadvantages of empirical soil-erosion models. To overcome these limitations, we introduce a new approach that integrates empirical and artificial intelligence models. Erosion-prone locations (erosion ≥16 tons/ha/year) are identified using RUSLE model and a soil-erosion map is prepared using random forest (RF), artificial neural network (ANN), classification tree analysis (CTA), and generalized linear model (GLM). This study uses 13 factors affecting soil erosion in the Talar watershed, Iran, to increase prediction accuracy. The results reveal that the RF model has the highest prediction performance (AUC=0.95, Kappa=0.87, Accuracy=0.93, and Bias=0.88), outperforming the three machine-learning models. The results show that slope angle, land use/land cover, elevation, and rainfall erosivity are the factors that contribute the most to soil erosion propensity in the watershed. Curvature and topography position index (TPI) were removed from the analysis due to multicollinearity with other factors. The results can be used to improve the identification of hot spots of soil erosion, especially in watersheds for which soil-erosion data are limited.