Exploring Bayesian network model with noise filtering for rainfall-induced landslide susceptibility assessment in Fujian, China

Abstract

Machine learning models have been increasingly popular in landslide susceptibility mapping based on the correlations among landslides and their inducing factors. However, mislabeled data in model training sets would deteriorate model accuracy. This study employed a Bayesian network to analyze influencing factors on landslides in Fujian Province, China, prone to typhoons and landslides. An inventory of 5,992 historical landslides informs Bayesian network modeling, with ten geoenvironmental factors as predictors. We introduced a progressive noise filtering method to mitigate the mislabeling effects of non-landslide points. The results show that altitude, wind speed, and lithology are the most important factors of landslides in the study area. The accuracy of the resultant landslide susceptibility map was verified using the area under the receiver operating characteristic curve (AUC) and Moran’s I index. The AUC value was improved from 0.838 to 0.931 during the progressive noise filtering. The correlation between historical landslide number density (LND) and resultant landslide susceptibility index (LSI) was evaluated. The Local Indicators of Spatial Association based on Moran’s I index shows consistent distribution patterns for high LND and high LSI regions. This study provides a useful reference for reliable landslide susceptibility mapping in the study area and similar areas.