Investigation of an end-to-end neural architecture for image-based source term estimation

Rapid and accurate estimation of hazardous material release parameters, including source location, release time, and quantity of material released, is crucial for protecting assets and facilitating timely and effective emergency response. In this paper, we present a first artificial neural network (ANN) approach for end-to-end source term estimation (STE) using time-series of multispectral satellite images. The architecture consists of two successive ANNs. The first-stage ANN estimates the hazardous material release rate over time, producing a 3D concentration map, while the second-stage ANN utilizes the generated concentration map to estimate the 2D source location, release time, and easterly and northerly wind speeds. By leveraging the inherent nonlinearity of ANNs and advances in parallel computing, our proposed method aims to eventually overcome the limitations of existing optimization and Bayesian inference techniques in handling the nonlinear STE problem. In this preliminary study, we validate the performance of our approach on a simulated dataset, demonstrating its potential for enhancing the accuracy and speed of STE in real-world applications.