Predicting the flood susceptibility under land use and climate change scenarios using deep learning algorithms

Flood risk in semi-arid, snow-fed basins is increasingly influenced by both land-use and climate change, yet their combined future effects remain poorly quantified. This study predicts future flood generation potential (FGP) under combined land-use and climate scenarios in the Gharesou Watershed (Iran) using deep learning. Future land-use (2034-2054) was simulated via Markov chain, and climate variables (temperature, precipitation) under three SSP scenarios were downscaled using the change-factor method. FGP was mapped using CNN, MLP, and DNN algorithms, validated against observed discharge data. By 2054, natural vegetation is projected to decline by 20.9% of the watershed area, while agricultural and residential lands expand. Temperature rises by 3.5-4.5 °C, and although annual maximum precipitation declines, extreme events become more frequent. Under the optimized CNN model, high- to very-high-risk zones expand from 62% to 87% of the watershed. This study provides the first quantitative attribution of future flood risk in a snow-fed semi-arid basin, identifying land-use change as the dominant driver (about60-70% of increased risk) and climate change as an intensifier (about 30-40%). These results indicate that protecting natural vegetation and restricting land-use conversion in high-risk zones are more urgent than climate adaptation alone. Proactive policies (restoring rangelands/forests, integrating climate scenarios into spatial planning, and enforcing land-use regulations) are essential to enhance watershed resilience.