Random forest-based prediction of flash flood susceptibility under land-use dynamics in East Java watersheds

Abstract

The frequent occurrence of flash floods in the small mountainous areas around Argopuro, Ijen and Raung in East Java, with steep terrain and short hydrological response times, poses significant challenges for risk assessment in regions with limited hydrological observations. This study assesses flash-flood susceptibility in eight small watersheds in East Java, Indonesia, using a random forest (RF)-based approach. A flash-flood inventory for 2012–2022 was compiled and used to train and validate the model. Ten conditioning factors were derived from a Digital Elevation Model (DEM), river-network data and satellite imagery, including slope, topographic wetness index (TWI), topographic position index (TPI), river density, land use and normalised difference vegetation index (NDVI). Land-use changes from 2001 to 2020 were used to assess their influence. To strengthen the hydrological context, event-based rainfall data corresponding to flash-flood occurrences and field-based photographic documentation were incorporated. The accuracy of the RF model was assessed using ROC curves and root mean square error (RMSE), indicating outstanding performance (scores of 0.96–0.99). Based on the model results, four main factors were identified as key determinants of flash-flood occurrence: land cover and its changes, slope steepness, river-network density and TPI. The spatial patterns generated by the RF model delineate areas with varying susceptibility levels, providing actionable information for prioritising mitigation measures. These findings contribute to a better understanding of flash-flood drivers in small watersheds and provide a scientific basis for local governments and stakeholders to design targeted risk-reduction strategies that combine land-use planning, watershed management and community-based adaptation.