Physics-Informed Stacking Machine Learning for Surface Methane Estimation via Spectral Radiance Channel Selection

Atmospheric methane is the second-largest contributor to radiative forcing among long-lived greenhouse gases, and characterizing its spatiotemporal distribution is essential for identifying sources and evaluating mitigation policies. Methane emissions are most directly reflected in near-surface concentrations, whereas satellite observations retrieve only the column-averaged methane concentrations. This study presents a physics-informed stacking ensemble framework that estimates surface CH4 directly from TROPOMI SWIR observations. SWIR radiance channels within the 2305–2385 nm band were optimally selected on the basis of their CH4 information content, including the contribution of the lower atmosphere, as quantified by radiative transfer analysis, and these L1B radiances were combined with L2 products and auxiliary variables governing surface concentration formation. Predictions from machine learning models with different learning schemes were fused through a stacking ensemble that exploited their complementary error structures and outperformed all individual base learners against ground-based in situ observations. For years and stations that were not used in training, the ensemble achieved R2 = 0.75 and 0.66, respectively. The selected channels improved estimation performance, and the L1B-only configuration performed comparably without a dedicated CH4 retrieval product, indicating that the observed spectra carry information relevant to surface CH4 before retrieval. The reconstructed global fields reproduced large-scale spatial and seasonal variability with stronger regional contrasts than the input column observations, demonstrating that physically informed channel selection combined with machine learning enables satellite-based surface CH4 estimation.