Banu Priya M.Dr. Rajesh Kumar Prasad2026-01-142026-01-142026https://ir.chanakyauniversity.edu.in/handle/123456789/182The increasing number and intensity of floods in India show the urgent need for better early-warning systems that combine different data sources and modeling methods. To address this need, this paper introduces the Multi-Source Hybrid Framework (MSHF). MSHF merges an LSTM-Transformer module for analyzing rainfall and discharge patterns over time, a Graph Neural Network for understanding spatial dependencies in river networks, a satellite-radar fusion pipeline for detailed precipitation mapping, and a social media sentiment analyzer for gathering situational awareness from crowdsourcing. All these elements are connected by an attention-based ensemble meta-learner. MSHF runs on a cloud platform that processes real-time gauge and weather data from over 200 stations in the Brahmaputra, Ganga, and Krishna basins. It generates rolling six hour forecasts with an end-to-end delay of under five minutes. Trained on historical events from the INDOFLOODS database and validated with data from 2021 to 2024, MSHF achieves over ninety-five percent predictive accuracy and high hydrological efficiency. A comparative evaluation against standalone LSTM-Transformer, GNN, satellite-radar, and social media models uses paired t-tests and chi-square goodness-of-fit tests to confirm significant statistical improvements. Targeted ablation experiments—Temporal, Spatial, Fusion, Social, and Attention Ablations—show that removing any component significantly reduces performance. Operational trials also demonstrate consistent detection of severe flood events with few false alarms across various lead times and river basins. This emphasizes MSHF’s readiness for proactive flood risk reduction in India’s diverse hydrological contexts.enFlood forecastingLSTM-TransformerGraph Neural NetworksINDOFLOODSMulti-source data fusionMachine learningtext::conference output::conference proceedings::conference paper