Soil erosion susceptibility assessment in the Bistrița River basin (Romania) using machine learning algorithms and GIS

This study evaluates the relationship between soil erosion and 12 environmental and anthropogenic variables to identify erosion-susceptible areas in the Bistrita River basin (Romania), using machine learning algorithms (MLA). Three supervised classification algorithms Support Vector Machine (SVM), K-Nearest Neighbors (K-NN) and Random Forests (RF) were trained using 4761 sets of values, 1191 values to validate the models, and 1488 to test the models and to ensure that they could be applied in practice. The performance of each model was evaluated using Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²). Using natural breaks, the results were spatially represented using 5 erosion susceptibility classes (very high, high, moderate, low, and very low). The best trained MLA (RF with R² = 0.67) was used to simulate three scenarios i.e., what happens in 2050, 1) if the current deforestation trend (2001-2023) is continued and precipitation is reduced, on average by 0.5 mm/y, 2) if the trend is reversed and the forested areas of 2001 are returned, or 3) if the sub-Carpathian area, the most affected by erosion, is afforested by > 50,000 ha. In this last scenario, the model showed that the areas with very high erosion (23.2% of the surface area) are transformed into the high and moderate classes, but the estimated costs of planting and maintaining the forested areas are estimated at approximately 775 million euros, which may represent a serious limitation in achieving this goal. In the Subcarpathian area, where erosion is substantial, simulations have shown that afforestation of at least 50 ha/km² significantly reduces the phenomenon. Each 10 ha increase above this value causes a reduction in erosion by 1 t/ha/year, so at 90 ha/km² an erosion of 1 t/ha/year is reached. Developing scenarios to assess susceptibility to erosion using machine learning (ML) training models with the variable of increasing/decreasing different environmental and anthropogenic variables could help local authorities or various administrators to improve their ecosystem management.