SVD-LSTM-based rainfall threshold prediction for rainfall-induced landslides in Chongqing

Abstract

Rainfall-induced landslides in Chongqing, a region of significant interest due to its high incidence rate, have traditionally been predicted using empirical rainfall thresholds. However, these approaches suffer from regional limitations and differing levels of accuracy. This paper presents a novel prediction method for rainfall thresholds, based on Singular Value Decomposition Long Short-Term Memory (SVD-LSTM) networks, applied to the case of 148 rainfall-induced landslides in Chongqing. By utilizing Singular Value Decomposition (SVD) to decompose Long Short-Term Memory (LSTM) layer weights into two smaller matrices and adding a custom layer to the standard LSTM structure, the SVD-LSTM method reduces the dimensionality of weights in the input and intermediate layers, reducing computational complexity and accelerating model training. This multi-layer grouping concept provides a method to improve the accuracy and efficiency of rainfall threshold prediction for landslides, offering a robust solution to the geographic limitations and accuracy discrepancies inherent in empirical models.

Keywords

• Rainfall-induced landslides
• rainfall threshold
• prediction model
• SVD-LSTM
• landslide prediction
• singular value Decomposition