Tracking Pollution Sources Across Watersheds via Causal Graph Discovery

Abstract

The accurate identification of pollution sources across complex watershed systems remains a fundamental challenge in environmental science. Traditional monitoring approaches relying on grab sampling and correlation statistics are frequently inadequate to capture the directional, time-lagged causal pathways through which contaminants propagate downstream. This paper proposes an integrated framework for tracking pollution sources across watersheds using causal graph discovery (CGD) techniques grounded in directed acyclic graphs (DAGs). The framework combines multivariate time-series observations from distributed water quality sensors with constraint-based and score-based structure learning algorithms, notably the Peter-Clark Momentary Conditional Independence (PCMCI) method and the Greedy Equivalence Search (GES) algorithm, to recover a sparse, interpretable causal network of pollutant transmission pathways. A structural equation model (SEM) is subsequently fitted to the recovered DAG to quantify causal effect strengths between upstream and downstream monitoring stations. Experimental validation on simulated watershed data demonstrates that the proposed approach correctly identifies the principal nonpoint source pollution (NPS) input nodes with a precision of 0.87 and recall of 0.82 under realistic observational noise conditions. Comparisons with Granger causality and transfer entropy (TE) baselines confirm the framework's superior robustness to confounded co-variation. The results underscore the transformative potential of causal graph methods for real-time, data-driven pollution source attribution in large, instrumented watershed systems.