Abstract

Accurately identifying pollution sources in river systems is critical for environmental monitoring and intervention design, yet remains challenging due to sparse observations and complex transport dynamics. This work considers the inverse problem of source apportionment in networked river systems, where multiple upstream configurations can produce similar downstream signals. We develop a probabilistic framework that combines physics-based transport models with statistical inference to recover key features of latent pollution sources, including their timing, location, and magnitude, while accounting for uncertainty in both model structure and parameters. Simulation studies show that key aspects of source behaviour can be recovered with suitable prior information, but performance is sensitive to sampling frequency, data sparsity, and model assumptions. We also explore how auxiliary high-frequency proxy measurements can improve estimation and illustrate the approach on data from a local catchment.