Methods and validation

How the estimates are made

1. Passes. Every time a satellite or aircraft looked at a landfill is a pass, whether or not it saw a plume. Passes with too little coverage or too many clouds are set aside.
2. Detections. Plumes inside the site outline or within 500 m are attributed to the site. Plumes near another methane source are marked ambiguous and never assigned.
3. The model. For each site, a statistical model estimates how often it was emitting and how much, allowing for each instrument's chance of missing a plume of a given size. Every estimate comes with a 90% range. An instrument's passes enter the model only once a plume search has run on them, because a pass nobody searched is not a pass that saw nothing; each site page names any instrument left out.
4. Checks. For each dated inspection, notice or repair, the estimate after the date is compared with what control sites suggest would have happened anyway. A drop or rise is only called when the probability is at least 97.5%; otherwise the result is inconclusive. The smallest drop the test could have caught is reported with it.
5. Operator comparison. The satellite estimate is set beside the annual methane the operator reported to the EPA.

Limits

• Sentinel-2 sees only plumes of about 2 tonnes per hour or more.
• Satellites see only in daytime and clear skies, and sources publish about 30 days after an observation.
• Annual figures are methane above a detection floor in those conditions, never a site's total.
• A change that followed a date is not shown to have been caused by it.

Clouds in Sentinel-2 images

Clouds and their shadows can look like plumes to the Sentinel-2 method and decide which passes count. How the cloud model tells them apart, and how well it did on images it never saw.

Method checks