Deep Learning-Guided Design of Optimal Sensor Networks for Accurate Pollutant Source Detection

Abstract

Efficient sensor network design for atmospheric pollution source localization remains a major challenge in environmental monitoring. Selecting optimal sensor locations from numerous candidates is difficult under complex spatial and atmospheric conditions. Traditional methods, including trajectory-based metaheuristics and evolutionary algorithms, often entail high computational costs and limited scalability. To overcome these limitations, we propose the prediction of optimal sensor configurations (POSCs), a data-driven framework that reformulates sensor placement as a prediction problem. POSC operates in two stages. First, large-scale Gaussian plume simulations are performed to generate concentration maps. Then, a genetic algorithm (GA) is applied under varying atmospheric conditions to obtain multiple candidate optimal sensor configurations. The resulting supervised dataset pairs concentration fields with their corresponding optimal networks. Second, the data are used to predict new probability maps of sensor relevance, providing a framework that highlights the most informative sensor locations across scenarios through a convolutional self-attention residual network (CSAResNet). The model is first evaluated on its ability to predict probability maps of optimal sensor locations, reaching an AUC of 0.95. From these maps, a subset of sensors is selected, enabling source detection performance comparable to evolutionary methods (GA) while providing real-time prediction at negligible computational cost.