Improving landslide triggering thresholds using artificial neural networks and reanalysis multi-layer soil moisture information

Landslide prediction is crucial for the design of early warning systems, and the integration of soil moisture information aims to enhance the accuracy of such predictions. This study focuses on the development of artificial neural networks (ANNs) designed to recognize conditions that trigger landslides, incorporating soil moisture data alongside precipitation. Specifically, using ANNs, we investigate the advantage of deriving thresholds without a specific parametric equation, and, due to their flexibility to incorporate multiple input variables, they allow for a comprehensive analysis of landslides. Specifically, the research utilizes observed precipitation and ERA5-Land reanalysis soil moisture data at four different depth layers. To assess the effectiveness of the proposed approach under diverse climatic and geomorphological conditions, two distinct case studies are considered, namely Sicily Island (Italy) and a group of catchments in the Bergen area of Norway. The proposed methodology involves three main steps: i) the acquisition of rainfall and landslide data; ii) the creation of a database of triggering (TE) and non-triggering (NTE) events; iii) the development of ANNs predicting when a landslide is triggered from input precipitation and soil moisture data. A measure of the prediction uncertainty of the developed ANN models, related to the fact that a limited sample of triggering events may be available, is also carried out. Overall, the developed ANN classifiers, incorporating soil moisture information in addition to precipitation, prove to have better predictive performance than those relying solely on precipitation data. In our study, we also carry out comparisons to traditional power law thresholds, derived by optimizing the true skill statistic (TSS) based on cumulative precipitation and duration (E-D). While the power law E-D thresholds reach a TSS equal to 0.50 for both study areas, the inclusion of soil moisture information can lead to significant performance improvements, yielding TSS values up to about 0.90. These results corroborate the potentialities of the use of soil moisture information and machine learning techniques in improving landslide prediction.