Soil erosion susceptibility assessment based on the hybrid deep learning models and SHapley analyses

Abstract

This study assesses spatial soil erosion susceptibility in Nghe An province, Vietnam, using four novel deep learning models: Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM architecture. A key innovation of this study is the integration of temporal sequence modeling with spatial feature learning via a hybrid CNN-LSTM architecture, an approach not commonly used in soil erosion studies. In addition, the SHapley Additive exPlanations (SHAP) technique was used to explain how the model worked and how it performed. CNN-LSTM demonstrated the highest validation accuracy (85.7%), sensitivity (0.964), Kappa (0.714), and area under the curve (AUC) of 0.927, along with the lowest error (0.378), outperforming all other models in prediction capability and robustness. While the LSTM model achieved the highest training AUC (0.947) and accuracy (87.7%), it showed signs of overfitting, with a validation accuracy of 83.9%. CNN offered stable performance (validation AUC: 0.892), while DNN lagged, limited by its lack of temporal context (validation AUC: 0.884). Using SHAP, elevation, geology, and aspect were identified as the most critical drivers of erosion susceptibility. The findings underscore the superiority of hybrid deep learning models in capturing the spatiotemporal complexity of erosion processes. The proposed CNN-LSTM model demonstrates strong potential for real-world implementation environments, where both spatial and temporal patterns influence erosion dynamics. This research presents a novel framework for erosion-susceptibility mapping and highlights the importance of integrating interpretable machine learning (via SHAP) with advanced modeling strategies.