Evidence note: This article rests on primary publications by the organisations described — two Google blog posts, a company post by Descartes Labs, and public code repositories — alongside general encyclopedia entries. None of the technical claims below were independently re-tested or benchmarked for this article, and no experiment was run. Where a source is the subject’s own account of its own work, that is stated. Individual claims carry their own evidence marker; the general reservation is not repeated section by section.
A methane map looks like one kind of object: a coloured layer over a basemap, hotter where the gas is thicker. It is usually at least two kinds of object stacked together. Part of it is a reading — photons counted by an instrument in orbit and inverted into a concentration. Part of it is a model — an estimate of what a facility of a given type, size and age probably emits, painted onto a coordinate. Both are legitimate. They fail in completely different ways, and a layer that mixes them without saying so is difficult to use responsibly. PlayAgit’s earlier overview of how AI and satellites map global methane emissions from space surveyed the players; this piece is about the seam between the two kinds of layer.
What Google Research Built: A Deep Learning Model That Maps Global Methane Emissions From Satellite Data
Google Research published a write-up titled "Mapping global methane emissions from space with deep learning," describing a machine learning approach to producing methane emissions information from satellite observations (research.google). That the post exists, and that it presents the work as Google’s own, is established by the post itself.
What the post does not settle, for a reader deciding whether to use such a layer, is the part that matters operationally: the model architecture, the training data, the spatial resolution of the output, the validation protocol, and any error bars against ground truth. (Evidence: the sources available here do not establish these specifics; treat any figure circulating secondhand as unverified.) Nothing in the material reviewed here supports quoting an accuracy number for this model. A public repository named for the same problem space exists (mapping-methane on GitHub), but its relationship to the Google work — whether it reimplements, extends, or merely shares a topic with it — is not established by these sources and should be checked before it is cited as an implementation of anything.
The Same Work on Two Google Blogs, and Why the Venue Changes How You Read It
The same subject appears on Google’s general blog under the title "Mapping global methane emissions from space" (blog.google), alongside the research-blog version. Two venues, one body of work.
This is worth a developer’s attention rather than a shrug. The research-blog channel is written for practitioners and is the one that ordinarily carries method detail, ablations and links to papers or code. The corporate-blog channel is written for a general and policy-adjacent audience and tends to foreground significance over method. When a claim about the work circulates — in a press summary, a procurement deck, a downstream dataset card — the useful question is which of the two versions it descends from. A performance claim that traces only to the general-audience telling has, in effect, no method behind it that a reader can inspect. (Evidence: this is editorial analysis of the two posts’ framing, not a claim sourced from either.)
The Instrument Underneath: TROPOMI Aboard the Copernicus Sentinel-5P Satellite
Most global methane concentration layers in public circulation are commonly attributed to TROPOMI, the spectrometer carried by the European Union’s Copernicus Sentinel-5P satellite. (Evidence: uncorroborated here — the relationship between TROPOMI data and global methane concentration maps was not verified against mission documentation for this article, and remains an open item.)
The distinction that survives regardless is between what a spectrometer of this class measures and what a user usually wants. An orbiting instrument measures a column concentration — roughly, how much gas sits in the air between the sensor and the ground, along that line of sight. An emissions figure — kilograms per hour leaving a specific pipe, pond or pit — is not measured. It is inferred, by combining the concentration field with wind and transport assumptions. Every step of that inference is a place where a number can be right in the aggregate and wrong at a single site. Sources consulted here do not quantify how large that gap is for any specific product.
Descartes Labs, Kayrros and the Satellite Data Reported Behind the COP26 Methane Reduction Pledge
Descartes Labs has a post on sharpening the focus on methane emissions (Descartes Labs on Medium). Whether that constitutes published work on global methane mapping, as distinct from regional or targeted detection, is an open question on the evidence gathered here. (Evidence: uncorroborated; the scope of the company’s methane work was not independently confirmed.)
Kayrros is described in reference material as a company working on methane analytics (Wikipedia). Reporting has connected the company to global-scale methane mapping and to the satellite data cited around the methane reduction pledge made at the Glasgow climate conference in November 2021. (Evidence: uncorroborated — neither the company’s role nor the specific provenance of the data cited at that conference was verified for this article; both remain open.) The pledge itself is widely reported as a collective commitment to cut methane emissions roughly 30 percent below 2020 levels by 2030, though that figure was not checked against a primary text here.
The methane-as-greenhouse-gas framing that motivates all of this is standard in reference material on human-caused emissions (Wikipedia), and is presented here as background rather than as a finding of this article.
Where the Map Stops Being a Reading: GeoWaste’s AI, Geospatial Analysis and Emissions Modelling
The GeoWaste database is published as a public repository that maps and characterises major dumpsites globally (GeoWaste on GitHub). Its own description of its method combines AI, geospatial analysis and methane emissions modelling. (Evidence: uncorroborated — the composition of the method, and the provenance of the emissions figures in the database, were not verified for this article and are open.)
That combination is exactly the seam. Site locations derived from imagery are, in principle, checkable against a photograph. Site emissions produced by a model are a function of assumptions — waste volume, composition, age, capture infrastructure — and cannot be checked against a photograph at all. A dumpsite layer can therefore be highly accurate about where and systematically off about how much, with nothing in the rendering to distinguish the two.
Observed Versus Modelled: How to Read Any Global Methane Layer
A workable triage, applicable to any methane product a developer is handed:
- Retrieval. A concentration field inverted from radiance. Traceable to an instrument, a date, an overpass. Fails through cloud, low albedo, high latitude and coarse pixels — and fails visibly, as gaps.
- Inversion. An emissions rate inferred from a concentration field plus meteorology. Traceable to an algorithm and a wind product. Fails through transport error, and fails quietly.
- Inventory or model. An emissions estimate assigned to a facility from its attributes. Traceable to an emission factor and an activity dataset. Has full global coverage by construction, including over places no instrument has usefully seen — which is precisely why it never looks like it is missing anything.
Coverage completeness is the tell. A layer with no holes is rarely a layer with no ignorance. (Evidence: editorial framework, offered as an analytic tool rather than a sourced result.)
Deep Learning at the Edges of the Same Problem: Recognizing Industrial Smoke From Imagery
Deep learning is generally described as a branch of machine learning built on multi-layer neural representations (Wikipedia), and similar techniques have reportedly been applied to recognising industrial smoke in video and imagery — one public example being the deep smoke machine project from CMU’s CREATE Lab (GitHub). (Evidence: uncorroborated; the project’s performance and current status were not assessed here, and both remain open.)
The adjacency is instructive. Smoke recognition classifies a visible plume against human-labelled footage, so it can be audited by eye. Methane is invisible in that sense, which removes the cheapest available check and pushes the entire burden of trust onto the validation protocol.
Developer Takeaway: What to Ask Before You Trust a Methane Layer in Your Own Pipeline
Before a methane layer enters a product, the answerable questions are:
- Is this cell observed or modelled? If the schema cannot express the difference, the layer is not ready for decisions about a single site.
- What is the native resolution, before any resampling? A smooth raster says nothing about the pixel it came from.
- What is the observation date and revisit interval? A "current" layer may be a composite spanning weeks.
- Where is the validation, and against what ground truth? Absent that, an accuracy claim is a marketing figure.
- What happens to the gaps? Interpolation across cloud gaps can manufacture a plume that no instrument saw.
- What is the licence and the citation requirement? Public repositories vary widely, and this matters before redistribution.
For attribution to a named operator or facility, none of the public global layers reviewed here were shown to be adequate on their own.
What Has Not Been Tested Here, and the Follow-Up That Would Settle It
No dataset was downloaded, no model was run, and no figure in this article was produced by measurement. Every technical claim above traces to a publisher’s own account of its own work, and the items flagged as uncorroborated are genuinely open, not merely cautiously worded.
The follow-up that would settle the central question is narrow and reproducible: take a bounded region, pull the satellite-derived concentration layer and the modelled dumpsite emissions for the same footprint and the same period, and compare them site by site — recording software versions, retrieval dates, the count of sites compared, and every case where the two disagree, including the disagreements that are inconvenient. That comparison would turn the observed-versus-modelled distinction from a reading framework into a measured result. It has not been done here.