Publication date: September 2, 2026
Google Research released the trained MAPL-EMIT model on September 1, together with an inference library, synthetic training data and a global methane-plume database. The model analyzes NASA EMIT hyperspectral satellite imagery to detect methane plumes, estimate their enhancement and locate likely sources.
This is a deployable specialist model release, not only a research paper. Google provides the trained model on Kaggle and code for inference.
What is MAPL-EMIT and what can it do?
MAPL-EMIT stands for Methane Analysis and Plume Localization with EMIT. It is a deep-learning framework for turning hyperspectral satellite measurements into maps of methane emissions and likely source locations.
The system uses a Swin-S vision transformer and handles three connected tasks: estimating methane enhancement, delineating the plume and locating its source. Google trained it with 3.6 million synthetic methane plumes injected into real EMIT scenes.
For environmental analytics teams, the practical change is scale. A trained model and inference code can automate work that otherwise depends heavily on expert review and matched-filter processing.
MAPL-EMIT performance and release artifacts
| Item | Google Research detail |
|---|---|
| Expert-annotated plume recall | 84% |
| Training data | 3.6 million synthetic plumes in real EMIT scenes |
| Model artifact | Trained model on Kaggle |
| Inference | Public GitHub library |
| Additional data | Global plume database and synthetic plume dataset |
Google reports 84% recall on expert-annotated methane plumes. Across roughly 1,100 EMIT granules, the researchers say MAPL-EMIT identified about 50% more plausible plumes than the comparison process. Google also reports that the system mapped 24 of the 25 highest-emitting landfills in its analysis.
These figures come from Google’s research evaluation. They are not independent benchmarks performed for this article, and the release is designed around EMIT hyperspectral data rather than every satellite sensor.
What is missing from the announcement?
Google explicitly notes that false positives remain a challenge, particularly over complex terrain. That limitation matters for governments, energy operators and climate-data teams because a detection can trigger expensive follow-up work.
The release also does not provide a simple commercial support or hosted-API offer. Teams can download the model and inference library, but production deployment, monitoring and adaptation remain their responsibility.
There is no directly comparable vendor table against other methane-detection models using a standardized operational benchmark. Buyers should therefore evaluate precision, recall and false-positive costs on their own geography and sensor workflow.
What this means
Earth-observation teams, methane-monitoring programs and climate analytics vendors should care most. The first useful test is to run the public inference stack on known facilities and compare detections against existing expert-reviewed workflows.
Verdict: MAPL-EMIT is worth testing now for teams already working with EMIT or similar hyperspectral data. The open artifacts make evaluation practical, but the false-positive caveat means the model should support expert review rather than replace it outright.
For more information, visit the official announcement of MAPL-EMIT on the Google Research website.
