Inverse identification of a passive scalar source of pollution in large areas

This paper presents a novel and computationally efficient framework for the inverse identification of atmospheric pollution sources, with a focus on real-time applicability in complex environments. The proposed method combines stochastic modeling of pollutant dispersion with Higher-Order Singular Value Decomposition (HoSVD) and machine learning-based surrogate modeling. A case study centered on the Singapore region demonstrates the method’s ability to accurately reconstruct pollution fields from sparse sensor data over a 900-hour simulation window. The approach leverages a multi-objective optimization strategy, using Gaussian Process Regression (notably, the regression task is carried out in a parametric space of relatively low dimensionality) to approximate parametric modes and the Levenberg–Marquardt algorithm for source localization. Results show that the localization error remains within a fraction of the spatial grid resolution, and the entire computational pipeline—from simulation to optimization—can be executed in under 30 min on standard desktop hardware. This efficiency enables frequent or continuous deployment in operational monitoring systems. The framework’s modularity allows for extension to higher-dimensional parametric spaces, making it suitable for a wide range of environmental monitoring applications.