Abstract
<title>Abstract</title> <p>The increasing frequency and severity of flooding in India have exposed the limitations of single-source hydrological forecasting and have created an urgent need for decision support tools that can be embedded directly into the operation of water infrastructure-reservoirs, drainage networks, and early-warning systems—rather than treated as standalone hydrological exercises. This paper presents the Multi-Source Hybrid Framework (MSHF), an ensemble forecasting and risk assessment platform designed to support flood-resilient water infrastructure planning and operations in Indian river basins. MSHF combines an LSTM-Transformer module for temporal rainfall-discharge dynamics, a Graph Attention Network (GAT) for spatial river-network dependencies, a satellite-radar fusion pipeline for precipitation mapping, and a social-media sentiment analyser for crowdsourced situational awareness, all coordinated by an attention-based meta-learner that dynamically weights the contribution of each stream. The framework was evaluated on a simulated dataset spanning 66 years (1959–2024) across 30 gauge stations in the Brahmaputra, Ganga, and Krishna basins under a temporally disjoint split protocol (train: 1959–2020, test: 2021–2024). On the hold-out test set, MSHF achieved a mean Nash-Sutcliffe efficiency (NSE) of 0.329 ± 0.042 and a Kling-Gupta efficiency (KGE-2012) of 0.284 ± 0.066, outperforming single-source baselines (Plain LSTM, GAT Only, and classical models) under Wilcoxon signed-rank and Diebold-Mariano tests. Ablation experiments confirmed that every branch contributed measurably to the overall fusion skill, with the temporal branch carrying the largest contribution and social-media crowdsourcing adding a secondary, masked-robust situational signal. The results position MSHF as a component of broader water-infrastructure risk management.</p>