Use of delayed ERA5-Land soil moisture products for improving landslide early warning

Abstract

Previous studies have demonstrated that incorporating soil moisture information in landslide-triggering thresholds can improve their predictive performance. ECMWF ERA5-Land reanalysis soil moisture information has proven to have this potential. However, ERA5-Land data are released with a latency of usually 5 d, which limits their immediate operational use in Landslide Early Warning Systems (LEWSs). In this study, we investigate whether delayed soil moisture data - ranging from 0 to 15 d prior to rainfall events - can still effectively inform landslide-triggering conditions. Specifically, we develop artificial neural networks (ANNs) trained on various delay times and evaluate how detection performances vary with increasing lag. We measure performances by ROC-based indices, such as the True Skill Statistic (TSS). Focusing on Sicily, Italy, our results show that even delayed soil moisture data consistently outperform models based solely on rainfall (TSS = 0.68 vs. 0.59). Notably, TSS reduces only marginally, from 0.78 with no delay to 0.72 with 5 d delay, and 0.67 with 15 d delay. This performance remains higher than that obtained using only soil moisture data (without precipitation and no delay, TSS = 0.53), as well as those achieved with a traditional power-law threshold based on rainfall intensity and duration (TSS = 0.50) and also through ANN model using rainfall intensity and duration (TSS = 0.59). These findings are, thus, promising for an operational use of ERA5-Land soil moisture products in LEWSs.