AI-based rainfall prediction model for debris flows

Abstract Debris flow prediction based on rainfall monitoring is important for early warning and disaster risk reduction. Taking a typical debris flow catchment (with debris flows occurring on average 6 times per year and with a watershed area of 19.47 km2) in central China as an example, we used a machine learning approach (17 machine learning algorithms were tested) and a time series data processing algorithm (141 features were extracted from rainfall data) to predict debris flow events in advance, based on continuous rainfall records from five rain gauges in the catchment. Using 367 rainfall events (46 rainfall events triggered debris flows, and 321 rainfall events did not) from December 2012 to April 2015, rainfall prediction models for debris flows were established, which can automatically assess in terms of probability whether any given rainfall event will trigger a debris flow. After model optimization, the Extra Trees (ETs) model showed the best performance. Under testing with 16,968 rolling rainfall series which simulated actual rainfall monitoring records, the model had no false alarms and missing alarms. Taking the event on May 11, 2012 as an example, the model correctly predicted the debris flow event 35 min in advance, i.e., before the debris flow reached the catchment outlet, which provided valuable time for early warning.