A hierarchical graph-based hybrid neural networks with a self-screening strategy for landslide susceptibility prediction in the spatial-frequency domain

Landslide susceptibility prediction (LSP) is a complex task with unresolved uncertainties, such as errors in sample classification and intricate relationships among environmental factors and spatial grid units. Additionally, the absence of interpretable black box models restricts the credibility and effectiveness of prediction models. To tackle these problems, an innovative interpretable deep learning model based on self-filtering graph convolutional networks and long short-term memory (SGCN-LSTM) is proposed. In the SGCN-LSTM, a self-screening strategy is employed to remove landslide/non-landslide samples with substantial errors that fall outside a defined threshold interval. Furthermore, SGCN-LSTM extracts nonlinear connections between environmental factors and long-range dependencies among grid units through spatial nodes and information gates. The Anyuan County in south China, with 2,655,972 grid units, 16,594 labeled, served as the study area. The LSP models used numeric inputs from the Frequency Ratios of 10 environmental factors in these spatial grid units. Results show that the accuracy and area AUC of the SGCN-LSTM achieve 92.38% and 0.9782, which are higher than those of one deep learning model cascade-parallel long short-term memory and conditional random fields (by 5.88% and 0.0305), and four machine learning models (by 12.44-20.34% and 0.0532-0.1909). This article delves into SGCN-LSTM 's evaluation results using the SHAP method, providing insights into the landslide development patterns and spatial heterogeneity of associated environmental factors in Anyuan County, with a global interpretability perspective. In conclusion, the SGCN-LSTM automatically screens erroneous samples, effectively extracts nonlinear features and spatial relationships from various environmental factors and delivers superior prediction accuracy and robustness for LSP.