A Physics‐Informed Two‐Stage Retrieval Method for EMIT Remote Sensing Methane Plumes

Abstract Hyperspectral satellite missions, particularly EMIT, have revolutionized methane point source monitoring. However, current operational retrieval pipelines typically rely on physically‐based matched filters followed by labor‐intensive manual verification to mitigate high false alarm rates caused by complex background interference. This dependence on human intervention restricts the scalability and efficiency required for rapid, automated global monitoring. To address these limitations, this paper proposes Plume‐Net, a novel Physics‐Informed Guided Two‐Stage Retrieval Method. Unlike conventional approaches, Plume‐Net decouples the task into a “global existence classification” stage and a “local pixel‐level regression” stage to maximize reliability without human‐in‐the‐loop. The core innovation lies in the integration of physical priors with deep learning: In the first stage, a Physics‐Informed Spectral Gating (PISG) module explicitly injects methane spectral absorption knowledge into the network, dynamically recalibrating channel weights to suppress background noise. In the second stage, to overcome the difficulties of regressing sparse, long‐tail plume distributions, we propose a Plume‐Focused Sampling Strategy and an Intensity‐Aware Pixel‐Level Weighted Loss function, forcing the network to prioritize rare but critical high‐concentration signals. Evaluated on real‐world EMIT observations, Plume‐Net achieves an ROC‐AUC of 0.961 in detection and a Mean Absolute Error (MAE) of 33.33 ppm·m in quantification. By significantly reducing false positives while maintaining high precision, this method offers a robust, fully automated alternative to current semi‐manual retrieval workflows.