Machine Learning–Based Systems for Early Warning of Rainfall-Induced Landslide

Abstract

Landslide disasters have inflicted incalculable losses on China’s national economy, as well as on lives and property. Notably, 90% of landslide disasters are directly induced by rainfall or have indirect associations with it. In Bazhong City, Sichuan Province, China, the proportion of rainfall-induced landslides accounts for more than 70% of all geological disasters in the region. Our research undertook a susceptibility analysis of multimodal landslide data in Bazhou District of Bazhong City, employing four distinct machine learning methods: decision trees (DTs), random forests (RFs), support vector machines (SVMs), and back-propagation neural networks (BPNNs). Additionally, data from the Tropical Rainfall Measuring Mission (TRMM) 3B42 precipitation product were utilized to develop a rainfall intensity-duration (I-D) model for the Bazhou District. The experimental results indicated that the BPNN achieved the highest overall classification accuracy, reaching 92.00%, which was 3.00% to 6.00% higher than those achieved by other algorithms. The kappa coefficient for BPNN was 0.84, surpassing other algorithms by 0.06 to 0.10. Furthermore, our results demonstrated that the rainfall I-D model had a prediction accuracy of 90.91% for rainfall-induced landslides. Finally, a probability quantification model for landslide triggering factors was established based on the previous two research results, aimed at meteorological warning. Comparisons with five recorded landslide events in 2009 revealed that the experimental outcomes of the meteorological early warning model aligned with the actual inspection results. Therefore, this model can serve as a reliable reference for issuing warnings about rainfall-induced landslides in Bazhou District.