MethaneGuard - Physics-Informed AI for Methane Leak Detection & Prioritization

Methane causes roughly a third of warming to date and traps 84× more heat than CO₂ over 20 years, yet it's invisible and dominated by a few super-emitters - so the bottleneck isn't just spotting a plume, it's knowing how big it is and which to fix first. MethaneGuard is an end-to-end pipeline that turns shortwave-infrared imagery into a ranked inspection worklist. It follows the state-of-the-art STARCOP approach - a physics + ML fusion, not a black-box CNN — and extends it with the operational layers STARCOP leaves out. A mag1c-style matched filter reduces the spectral cube to a methane-enhancement product; a lightweight U-Net (MobileNetV2, 6.6M params) segments plumes from a 4-channel input (three context bands + the filter product), which rejects the roof/road/panel false positives a filter alone flags; Integrated Mass Enhancement converts each mask into an emission rate in kg/hr with CO₂e and uncertainty; and a risk engine scores every leak by emission, confidence, infrastructure proximity and persistence to produce a dispatch queue. A clean, map-centric dashboard - backed by a Flask server that runs the whole pipeline live on uploaded tiles — turns it into a working prototype that goes from satellite pixel to work order.

What I learned

That in this domain physics and ML are partners, not alternatives: a matched filter finds plumes but flags every bright surface, a CNN alone drowns on <1%-positive pixels, and feeding the physics product in as a channel and a loss weight is what makes detection reliable. I also learned that detection is the easy 60% - the operator value is in quantification and prioritization, so that's where I put the innovation. Hands-on, I got into SWIR absorption physics, the IME emission method, handling severe class imbalance with Dice + weighted BCE, and a sensor-agnostic ingestion path spanning AVIRIS-NG, EMIT and Sentinel.