Integrated Remote Sensing and Deep Learning Models for Flash Flood Detection Based on Spatio-temporal Land Use and Cover Changes in the Mediterranean Region

Rapid climate change is amplifying the frequency and severity of global flooding events. These floods induce declines in agricultural areas, water bodies, barren lands, precipitating diminished crop productivity due to habitat loss and constrained water availability. Conversely, urban sprawl, notably within high-risk flood zones, exhibits substantial expansion. Projections anticipated approximately 5200 km² of urban areas to confront heightened vulnerability to flash floods by 2030 and 2040, accentuating the exigency for immediate risk mitigation measures. This study scrutinizes the ramifications of flash floods on land use and land cover (LULC) dynamics over 20-years period within the Hodna watershed, situated in northern Algeria. The applied methodology integrates a random forest (RF) model for classification, complemented by a fused Cellular Automaton–Markov model to forecast future LULC trends for 2030 and 2040 based on the remote sensing data obtained from Landsat 5 and 8. The modeling results attained high prediction accuracy (Kno: 0.7857, Klocation: 0.8184, Kstandard: 0.7763), affirming the proposed methodology reliability. In addition, the study explored the employment of convolutional neural network (CNN) model coupled with Geographic Information Systems (GIS) for flood susceptibility and was delineated with 89% accuracy. The findings underscore significant susceptibility to flash floods, driven by hydrological and topographic factors, and explained through principal conditioning factors.