Real-Time Pollution Source Identification for Watershed Emergency Management Using a Multimodal Spatiotemporal Synergy Network

Abstract Accurate and timely identification of pollution sources is the core of water-pollution emergency management. We propose a multimodal spatiotemporal synergy network that integrates Internet of Things sensor networks, Unmanned Aerial Vehicle aerial imagery, and Geographic Information System data within a Convolutional Neural Network-Gated Recurrent Unit-squeeze-and-excitation-You Only Look Once version 5 framework capable of delivering command-ready information within 10 h after a pollution event. The system simultaneously ingests hourly sensor data on water-quality parameters and 0.3 m resolution aerial images; an adaptive attention module recalibrates channel-wise weights, cutting the mean absolute error of pollutant-concentration prediction to 0.008 mg L − 1 and confining pollution-source localization error to within 300 m. A web-based visualization interface outputs ranked pollution hotspots and targeted inspection lists whose spatial agreement with authority-verified source locations reaches 96%. Diffusion-path reconstruction accuracy improves by 18.3% over conventional Long Short-Term Memory, and decision-response time is shortened from the usual 30 h to 10 h, a 67% reduction. The framework can be transferred to other basins with minor parameter tuning, strengthening timely response and adaptive management of sudden pollution incidents.