2026 workshop program
Plume Segmentation from MethaneSAT with Cross-Sensor Transfer Learning and Physics-Informed Postprocessing
The work
Abstract
Automated detection and masking of individual methane plumes from satellite imagery is essential for operational emission attribution and quantification. We present a machine learning framework for instance segmentation of methane plumes from MethaneSAT retrieved XCH_4 maps, addressing two core challenges: scarcity of labeled MethaneSAT data and the need for reliable inference across diverse atmospheric and surface conditions. We show that Mask R-CNN with a ResNet-50 backbone outperforms U-Net semantic segmentation on both MethaneAIR and MethaneSAT data, and that fine-tuning from MethaneAIR pre-trained weights is the most effective cross-sensor transfer strategy, achieving instance-level precision of 0.60 and recall of 0.98. A physics-informed post-processing pipeline produces two operational modes: a high-sensitivity mode (precision 0.71, recall 0.94) for comprehensive emission screening, and a high-reliability mode (precision 0.92, recall 0.70) for confident source attribution.