Landslide Susceptibility Mapping Using an LSTM Model with Feature-Selecting for the Yangtze River Basin in China

Abstract

Landslide susceptibility mapping (LSM) is crucial for disaster prevention in large, complex regions characterized by high-dimensional data. This study proposes a Feature-Selecting Long Short-Term Memory (FS-LSTM) framework to enhance LSM accuracy by integrating feature selection techniques with sequence-based modeling. The Mean Decrease Impurity (MDI) and Information Gain Ratio (IGR) were used to rank landslide conditioning factors (LCFs), and these rankings structured FS-LSTM inputs to assess the impact of feature ordering on model performance. Feature-ordering experiments demonstrated that structured rankings significantly improve model accuracy compared to randomized inputs. Our model outperformed traditional machine learning algorithms, such as logistic regression and Support Vector Machine, as well as standard deep learning models like CNN and basic LSTM, achieving a score of 0.988. The MDI and IGR rankings consistently identified soil type, elevation, and average annual cumulated rainfall as the most influential LCFs, improving the interpretability of the results. Applied to the Yangtze River Basin, the FS-LSTM framework effectively identified landslide-prone areas, aligning with known geological patterns. These findings highlight the potential of combining feature selection with sequence-sensitive deep learning to enhance the robustness and interpretability of LSM. Future studies could expand this approach to other regions and incorporate real-time monitoring systems for dynamic disaster management.