A hybrid approach for enhanced flood prediction and assessment: Leveraging physical models, deep learning and satellite remote sensing
Mohammad Roohi, Hamid Reza Ghafouri, Seyed Mohammad Ashrafi, Mahdi Motagh, Mahmud Haghshenas Haghighi
Abstract
Accurate real-time information is crucial for effective flood risk management, especially in regions with complex terrain and irregular rainfall patterns. This study developed a hybrid model integrating process-based hydrological modeling, ensemble learning, and deep learning to enhance flood prediction and floodplain assessment in the Dez Basin, Iran, a flood-prone region. The proposed framework couples the HEC-HMS hydrological model with ensemble learning algorithms, employing base learners to train the ensemble model and combining their outputs via a neural network. This approach significantly improved prediction accuracy, achieving R2 values of 0.81–0.88, thereby addressing the challenges of precise flow prediction in complex regions. Simultaneously, the U-Net model processed 342 Sentinel-2, Landsat 8, and Landsat 9 images from Google Earth Engine, leveraging NDWI, MNDWI, and NDMI spectral indices to delineate flood extents for six flood events (April 2016, January and April 2019, and March–May 2023) with a mean Intersection over Union (mIoU) of 70–71.3%. Though not quantitatively linked due to computational constraints, the alignment of peak flows with mapped extents bridges temporal and spatial flood analysis, enhancing real-time management. This scalable framework, through accurate peak flow predictions from the hybrid model and precise flood extent mapping via U-Net, strengthens real-time early warning systems and supports targeted disaster preparedness, offering actionable strategies for global flood-prone regions to mitigate future risks.