Plume Segmentation from MethaneSAT with Cross-Sensor Transfer Learning and Physics-Informed Postprocessing
Researchers have developed a machine learning framework to detect methane plumes from satellite imagery, specifically addressing challenges with limited labeled data from MethaneSAT. The system utilizes a Mask R-CNN model with a ResNet-50 backbone, outperforming U-Net and showing strong performance with cross-sensor transfer learning from MethaneAIR data. A physics-informed post-processing pipeline enhances reliability, offering both high-sensitivity and high-precision modes for emission screening and source attribution. AI
IMPACT Enhances capabilities for environmental monitoring and emission attribution using AI.