Physics-informed spatiotemporal analysis of methane concentrations in an oil sands region

Methane (CH 4 ) emissions from complex industrial regions, especially oil sands ponds, exhibit strong spatial heterogeneity and episodic extremes, posing persistent challenges for reliable regional-scale concentration estimation. While recent data-driven models have improved predictive accuracy, their physical consistency, robustness to observation sparsity, and behaviour under extreme conditions remain insufficiently examined. Here, we develop a multi-source, physics-informed, and spatially structured prediction framework that integrates ground-based monitoring stations with satellite-derived background information to estimate regional CH 4 concentrations. The proposed framework explicitly enforces transport-consistent structure and spatial regularity while remaining robust to data gaps and sensor failures. Across independent validation experiments, the model achieves coefficients of determination exceeding 0.80 and demonstrates stable performance under simulated station dropout rates of up to 50%. Uncertainty calibration shows near-nominal predictive interval coverage (0.916 for a nominal 95% interval) with a mean interval width of ±215 ppb, indicating reliable probabilistic characterization. Spatial diagnostics further confirm physically realistic smoothness, with a roughness ratio of 0.64 relative to baseline interpolation methods, and strong alignment with wind-resolved transport patterns (Spearman ρ = 0.817). Using a minimal post-hoc monotonic correction as a diagnostic tool, the recoverability of extreme concentration amplitudes can be quantitatively assessed without modifying the underlying predictive model. Overall, this study demonstrates that integrating physical, spatial, and observational constraints enables not only accurate but also scientifically interpretable and deployment-ready CH 4 predictions in complex industrial environments. • PINN model fuses ground and satellite data for regional methane monitoring. • Physics constraints ensure realistic pollutant transport patterns. • The framework reconstructs extreme emission events missed by sensors. • Robust predictions maintained even with 50% station data loss.