A proposal for an approach to mapping susceptibility to landslides using natural language processing and machine learning

Compiling an inventory is a fundamental step for carrying out assessments of landslide hazards. However, data in sufficient quantity and quality are not always available. Thus, this study puts forward an approach for drawing up a landslide inventory using textual data from telephone records, and for mapping hazards of landslides in an urban area. Forty thousand seven hundred ninety-two textual records and the naive Bayes algorithm were used to classify them, and these form the landslide inventory. After creating the inventory, the random forest algorithm with 12 conditioning variables was used to map landslide hazards. The text classification model obtained an accuracy of 0.8671 and a Kappa index of 0.8038. The hazard mapping model obtained accuracy of 0.9503 and an AUC (area under the curve)-ROC (receiver operating characteristics) of 0.9870. The results produced by the model were also compared with real landslides reported in news reports and were shown to be close to what had happened, thus demonstrating the ability of the proposed approach to predict landslides. Finally, the proposed approach can be used in simulation environments, thereby supporting strategic decision-making associated with hazard analysis.