AI-Driven River Morphology Mapping for Flood Risk and Sediment Dynamics in the Brahmaputra River,Eastern Himalaya

Climate change is accelerating fluvial hazards across high mountain regions, where river morphology critically influences flood risk, sediment transport dynamics, and broader landscape evolution. In this study, we develop and evaluate a comparative deep learning framework designed to automate river morphology mapping by integrating multimodal remote sensing data, specifically Sentinel-1 SAR and Sentinel-2 optical imagery across geomorphologically diverse reaches of the Brahmaputra River. We benchmarked three architectures : Attention U-Net, SegFormer, and a novel hybrid Transformer U-Net,for multi-class segmentation of river channels, mid-channel bars, and background terrain. To simulate realistic operational conditions, we generated weakly supervised training labels using spectral indices and unsupervised clustering in Google Earth Engine(GEE). We assessed model performance using the Dice coefficient, mean Intersection over Union (mIoU), and Boundary IoU (BIoU) as our primary evaluation metrics. Our hybrid Transformer U-Net demonstrated the strongest generalization capacity across previously unseen river reaches (Dice = 0.95–0.96; mIoU = 0.91–0.92), while also showing notably improved boundary precision for both morphological features (Bar BIoU = 0.49; River BIoU = 0.69). To demonstrate the practical applicability of our approach, we conducted a targeted case study on a particularly flood-prone reach of the Brahmaputra, focusing on planform morphological assessment. This analysis highlighted how effectively the model captures dynamic channel–bar transitions and identifies potential erosion risk zones. By combining rigorous technical benchmarking with practical geomorphological analysis, our work illustrates the broader potential of deep learning tools to support climate-resilient river management strategies, inform sediment planning decisions, and enhance hazard mitigation efforts in vulnerable Himalayan landscapes.