Exploring the potential of lagged ERA5-Land Soil Moisture Data for Real-Time Landslide Prediction Using Neural Networks

The identification of landslide triggering conditions is a fundamental step for the development of effective landslide early warning systems (EWSs), essential for reducing the risks and impacts of these natural disasters. Enhancing the predictive accuracy of these systems requires advanced methodologies such as artificial neural networks (ANNs) that can dynamically assess landslide triggering conditions. Recent advancements have demonstrated significant improvements in landslide prediction when using ANNs fed with observed precipitation and multilayered soil moisture data from ERA5-Land at the onset of rainfall events. However, ERA5-Land data are typically available with a delay of approximately five days, making their direct application to real-time prediction systems challenging. This study investigates the feasibility of utilizing lagged ERA5-Land soil moisture data for real-time landslide prediction and evaluates impacts on predictive performance. Neural networks were developed using soil moisture data lagged by 0 to 15 days prior to rainfall events. The test-application focused on the case study of Sicily, Italy, and revealed that lagged soil moisture data affect prediction accuracy, which still significantly higher than using just precipitation data. For the lags of interest, the reduction of performance is modest. Specifically, with a 5-day lag, the True Skill Statistic index decreased only marginally, from 0.78 to 0.72. These findings highlight the potential for incorporating ERA5-Land multilayered soil moisture data into operational LEWSs, even when using lagged datasets, with potential real-time applications.