Data-driven precipitation thresholds for landslides acceleration based on long-term in-situ monitoring (Italian Alps and Apennines)

This is the authors' abstract. We don't add key points for this paper.

Rainfall thresholds are commonly adopted for landslide forecasting; however, most existing formulations are derived from inventories of landslide occurrence and are not specifically designed to anticipate kinematic changes in already active deep-seated landslides. This study introduces a rainfall intensity–duration (ID) threshold framework explicitly linked to observed acceleration phases of active deep-seated landslides, using long-term displacement monitoring data. The analysis considers 12 monitored landslides located in the Emilia-Romagna Apennines and the eastern Alps (Italy), with monitoring records spanning 2008–2024. Acceleration onsets were identified using two complementary strategies: a manual, expert-based analysis of cumulative displacement time series and an automated procedure based on statistically significant velocity anomalies. Rainfall forcing was characterised using both total rainfall and a net rainfall proxy that accounts for evapotranspiration and snow accumulation and melt. For each approach, rainfall events associated and not associated with acceleration were reconstructed, and corresponding ID thresholds were derived. Threshold performance was evaluated using Receiver Operating Characteristic and Precision–Recall spaces, together with imbalance-aware verification indices. The results show that rainfall thresholds based on landslide acceleration reliably separate hydrological conditions leading to increased displacement rates from non-triggering rainfall. Thresholds derived from manual and automated acceleration detection approaches are structurally consistent, indicating robustness with respect to detection strategy. Net-rainfall thresholds systematically outperform total-rainfall thresholds and selected thresholds from the literature, particularly for long-duration rainfall events. The proposed framework offers a practical and engineering-oriented tool for site-specific early-warning systems on active deep-seated landslides, supporting the prediction of hazardous acceleration phases rather than landslide initiation.