Towards the Development of Machine-Learning-Based LEWS for Data-Scarce Environments

Given the increase in the frequency of disasters caused by landslides due to extreme precipitation events and unplanned urbanization, landslide early warning systems (LEWS) have been shown to be increasingly necessary as effective and cost-beneficial risk-reduction and damage mitigation tools. Recently, the use of machine learning techniques in the prediction of landslide triggering for application in LEWS has shown promise, with several examples in the literature demonstrating good results. However, the need for large volumes of data for training models for this purpose is a considerable obstacle to their broader application, especially in regions that lack good landslide inventories. This study tests the use of civil defense service records related to landslides as a proxy for the actual triggering of landslides, since peaks in the number of service calls are observed during such events. Supervised machine learning models (Support Vector Machine, Multilayer Perceptron, and Random Forest) were used and their performances were compared with that of a LEWS based on empirical thresholds already in operation in the study area. To this end, records from the Civil Defense of Petrópolis, Rio de Janeiro, Brazil, regarding occurrences registered during the period from 2015 to 2019 were obtained, as well as a historical precipitation series from a rain gauge situated in the same municipality with the same temporal coverage. After data processing, which removed spurious readings in both data sources, an input was created whose features consisted of precipitation accumulations and maximum intensities recorded in different temporal windows and whose label was the presence or absence of civil defense records on the same date. The results confirm the potential of using machine learning algorithms in LEWS, since the models based on Random Forest and Multilayer Perceptron presented Recall, F1-Score, and Balanced Accuracy considerably superior to those of the LEWS operating in the municipality. They indicated, as well, the need for improvements to the empirical thresholds used in Petrópolis, particularly the ones for the activation of warning sirens, whose activations were concentrated in only 2 of the 4 thresholds in use.