A method for automatic detection of methane plumes in MethaneSAT satellite images is proposed, based on Mask R-CNN with ResNet-50. The model outperforms U-Net in pixel-level F1-score by 10.49 for MethaneAIR and 5.48 for MethaneSAT. To overcome the scarcity of labeled satellite data, fine-tuning was performed on MethaneAIR aircraft data and synthetic plumes. The final instance-level precision was 0.60 with a recall of 0.98. Physics-informed post-processing yields two modes: a high-sensitivity mode (precision 0.71, recall 0.94) for screening, and a high-precision mode (precision 0.92, recall 0.70) for source attribution. Analysis of false positives revealed that many correspond to actual emissions missed by conservative labeling, meaning the reported metrics are a lower bound on performance.
The MethaneSAT satellite scans for leaks of метана — a gas that's invisible but packs a powerful planetary warming punch. Its "nose" is a спектрометр, a device that spots the subtlest color changes in light passing through the atmosphere — the telltale sign of methane.
But picking up a scent isn't enough — you have to single it out from thousands of others. Where humans once pored over images, now the program is trained like a search dog: first it lunges at anything suspicious, like a pup following every random trail, then a second algorithm — the "handler" — confirms the catch. The real surprise? When the bloodhound "barks" at an empty spot, it often turns out to be an actual leak that cautious human experts missed.
This kind of hunt lets us quickly repair pipelines and landfills, cutting methane emissions that trap 80 times more heat than углекислого газа over the first 20 years.
🎯 Over 20 years, methane traps 80 times more heat than carbon dioxide.
🎬 Someday, such orbital bloodhounds will sniff the atmospheres of exoplanets in search of life.