Gully erosion susceptibility mapping using the stacking ensemble machine learning method

Soil erosion poses a worldwide challenge, significantly impacting water and soil resources. Gully erosion, in particular, creates large channels that are difficult to remediate using conventional agricultural equipment, leading to substantial socio-economic losses. Identifying areas vulnerable to gully erosion is crucial for devising preventive action plans that reduce gully formation and associated damages, ultimately contributing to the achievement of sustainable development goals. However, the environmental factors linked to gully development, such as weather, vegetation, land management, and slope, are complex and vary across time and space. This research explores the potential of using a stacking ensemble learning technique for predicting gully erosion susceptibility. A dataset comprising 1000 gully erosion locations and 38 environmental factors in agricultural land in Jefferson County, Illinois, U.S., was employed to develop the stacking model for gully prediction. Collinearity detection involved calculating the correlation coefficient and Cramer's V, followed by feature selection based on mutual information. The chosen features were divided into training, validation, and test sets to establish a robust model and facilitate a more generalized performance evaluation. Random Forest (RF), K-Nearest Neighbors (KNN), Gradient Boosting Machines (GBM), and Deep Neural Networks (DNN) were employed as base models. The optimization of each base model was achieved through a combination of K-fold cross-validation and Bayesian optimization. Although RF, KNN, GBM, and DNN models demonstrated comparable performance, the DNN emerged as the best meta-model, exhibiting the highest accuracy. Furthermore, a comparison between base models and meta-models revealed that the stacking-based ensemble model outperformed single algorithms in terms of predictive accuracy. The findings of this study indicate that the stacking ensemble machine learning method holds promise for predicting gully erosion susceptibility in the study area. This information can aid policymakers and land managers in concentrating their prevention or mitigation efforts and allocating resources where they are most needed.