GIS Based Soil Erosion Susceptibility Assessment Using Deep Learning Models: A Case Study in the Mountainous Region of Nghe An, Vietnam

Abstract

In this study, the main objective is to evaluate soil erosion susceptibility in the mountainous region of Nghe An Province using three deep learning models: Long Short-Term Memory (LSTM), Deep Neural Network (DNN), and Deep Attention Network (DaNet). A total of 685 erosion points were identified from field surveys and satellite imagery, and nine spatial conditioning factors were used, including slope, aspect, curvature, elevation, rainfall, Normalized Difference Vegetation Index (NDVI), soil type, distance to faults, and geology. The dataset was split into 70% for training and 30% for validating. Various validation metrics including area under the ROC curve (AUC) were used for validation and comparison of the models. The results show that among the tested models, DaNet showed the highest predictive performance, achieving an AUC of 0.936 on the training dataset and 0.852 on the validating dataset compared with other deep learning models (DNN and LSTM). The susceptibility map produced by DaNet demonstrated strong spatial alignment with real-world erosion occurrences, with 59.88% of observed erosion points located in the very high susceptibility class and 18.52% in the high class, totaling 78.4% of all erosion events. These results confirm DaNet’s effectiveness in capturing complex spatial patterns and delivering reliable erosion risk predictions, supporting its use for land-use planning.