In dense oil and gas basins, hundreds of moderate methane leaks sit closer together than one satellite pixel. Their plumes blur into a single regional haze. Current methods can measure the basin total, or find the rare huge leak, but they cannot say which facilities are emitting. Muscle physiology has solved a structurally similar problem. A skin electrode records the overlapping electrical activity of many motor units, the bundles of muscle fibres each controlled by one nerve cell. Electromyography (EMG) decomposition recovers each unit's firing times because every unit has a fixed physical identity and fires sparsely and repeatedly. We propose to treat an oil basin the same way. Each facility is a persistent "source unit" with fixed physical parameters (position, release height) that switches on and off. Each satellite overpass records these sources mixed through that day's wind. Fitting one shared set of source parameters across many overpasses should separate sources that no single overpass can resolve, because the wind keeps changing direction.
The opportunity matters because heavy-tailed, intermittent emissions dominate oil and gas methane totals. Global assessment of oil and gas methane ultra-emitters used the TROPOMI satellite instrument with HYSPLIT dispersion modeling to find about 1,200 ultra-emitters above roughly 25 t/h. These account for 8–12% of global production emissions. The same study also found that high-emitting basins such as the Permian showed fewer resolvable ultra-emitters than expected, because their plumes merged. Below that detection threshold, attribution is currently blind.
Core Claim
Day-to-day wind variability plays the role in atmospheric sensing that a dense electrode grid plays in EMG: it supplies the diversity that makes a mixture separable. On any one day, two leaks 3 km apart are indistinguishable at TROPOMI's 5–7 km pixel scale. Over fifty days, though, their plumes point in different directions on different days, and so they leave different footprints. We claim that a joint model can recover individual sources at spacings below the pixel size and at rates well under the 25 t/h single-overpass threshold, without being told where the infrastructure is. The joint model has three parts:
- Persistent, off-grid source objects, each with a continuous location and release parameters.
- Sparse, intermittent emission time series, one per source.
- A differentiable transport model conditioned on each overpass's meteorology.
What is new is the persistent off-grid source object fit across time. Neither sparsity alone nor learned transport alone is new.
Where the Method Comes From and How It Differs
The decomposition machinery comes from two EMG papers.
Tutorial on MUedit supplies the classical pipeline. It separates channels by maximizing how sparse the recovered sources are, and it "peels off" sources one at a time: it reconstructs each identified unit's footprint, subtracts it, and searches the remainder for weaker units. The same tutorial also constrains us. It states that this pipeline requires a stationary mixing filter. The atmosphere violates that requirement every hour, so we cannot copy blind unmixing with one fixed unmixing matrix.
A Biophysical-Model-Informed Source Separation Framework For EMG Decomposition shows the way around this limitation. Its method, BMISS, replaces the arbitrary unmixing matrix with a neural surrogate of the physics of signal spread through tissue, conditioned on six anatomical parameters. It then optimizes those parameters by gradient descent to make the recovered sources as sparse as possible. On a synthetic 100-unit pool it recovered 86 units, against 5 for FastICA, the standard statistical unmixing algorithm. BMISS was tested only in simulation and degraded at an SNR of 5, and both caveats carry over to our setting.
Our setting differs from BMISS in one decisive way. BMISS's tissue filter is static, whereas our transport operator changes every overpass but is largely known from meteorology. The persistent parameters are therefore the source positions, not the filter. The inversion is a semi-blind problem, as opposed to the blind one EMG faces.
On the atmospheric side, standard Bayesian inversion balances fit to observations against a Gaussian prior centred on inventories. Definitions and methods to estimate regional land carbon fluxes for the second phase of the REgional Carbon Cycle Assessment and Processes Project (RECCAP-2) and Mesoscale inversion: first results from the CERES campaign with synthetic data both describe this setup. It provides the observation-error weighting we keep. Its Gaussian smoothing, however, is what spreads sparse sources across a basin, which is why we replace it.
The ultra-emitter study inverts one overpass and one isolated plume at a time. AI for operational methane emitter monitoring from space (MARS-S2L) detects plumes in high-resolution Sentinel-2 and Landsat imagery, but only at 500 kg/h and above, and it infers neither source time series nor overlapping mixtures. We use MARS-S2L as independent ground truth rather than as a competitor.
Research Program
The forward model predicts the methane enhancement in pixel $j$ on overpass $t$ as
$$y_{jt} = \sum_{k} q_{kt}\, g(\theta_k, m_t)_j + b_{jt} + \epsilon_{jt}.$$
Here $\theta_k$ is the continuous location and release height of source $k$ and is shared across all overpasses. The emission rate $q_{kt} \ge 0$ is that source's rate on overpass $t$. The footprint $g(\theta_k, m_t)$ is the plume pattern a unit release from $\theta_k$ makes under that overpass's meteorology $m_t$, taken from ERA5 winds and boundary-layer height. The term $b_{jt}$ is a smooth background, and $\epsilon_{jt}$ is retrieval noise.
We fit the model by minimizing the error-weighted residual. A penalty on the time series $q_{k\cdot}$ favours sources that are sparse and intermittent, playing the role kurtosis plays in EMG. The background $b$ is regularized to be smooth, using the same Gaussian, covariance-based regularization the inversion literature uses for its flux prior, so that diffuse sources (wetlands, distributed leakage) are absorbed there rather than hallucinated as point sources.
Sources enter one at a time by peel-off. We initialize at the largest residual hotspot aggregated across overpasses, fit that source, subtract its footprint, and repeat. A source is kept only if it reduces the held-out-overpass residual.
The footprint model is built in two stages. Stage one uses an analytic Gaussian plume, which is already differentiable. Stage two uses a neural surrogate trained on HYSPLIT or FLEXPART runs across meteorological conditions. This mirrors BioMime but is conditioned on changing inputs. Whether the surrogate earns its cost is itself an experiment.
To handle transport error, we let each overpass carry a small learned wind-direction and speed correction shared by all sources. This addresses the mismatch between simulated and real transport that BMISS left untested.
Baselines
The naive reference is the basin total from mass balance, with emissions spread uniformly or in proportion to well density. It gives the right total and no attribution, so it fixes the floor.
The standard method is a Gaussian-prior Bayesian inversion of the same cost function as in RECCAP-2 and CERES. It runs on a 1 km grid using all overpasses jointly, with the same footprint model, and uses no infrastructure locations so that the comparison stays fair.
The per-overpass point-source approach follows the ultra-emitter study. It fits HYSPLIT-style single-plume inversions to each overpass separately and then clusters the detections. This is the strongest existing published practice for attribution. Its reported operating point is a detection limit of about 25 t/h; no reported result found on Lacuna for its source-separation accuracy in dense clusters.
The strongest competitor we construct is a group-sparse grid inversion. It applies a penalty that makes each 1 km cell either active or inactive across all overpasses, with sparse time series inside active cells. It shares our persistence and sparsity assumptions and differs only in placing sources on a fixed grid without release-height parameters.
All methods get identical observations, meteorology, background model and a matched hyperparameter search budget, and each is tuned on separate synthetic basins. We beat the competitors if we recover more true sources at a matched false-discovery rate, and if we do so specifically in the regime where neighbouring sources sit closer than the pixel size.
Three ablations isolate the ingredients. The first removes temporal sparsity, leaving only persistent sources. The second removes persistence, fitting each overpass independently. The third swaps the neural surrogate for the analytic plume.
Why not just…?
Why not just run a high-resolution Gaussian inversion and wait for finer satellites? Gaussian priors penalize large departures quadratically. For emissions dominated by a few episodic events, that spreads each spike over its neighbours, as the ultra-emitter study's heavy-tailed distribution implies. Finer instruments such as Sentinel-2 help, but MARS-S2L shows they still see only sources above about 500 kg/h, and only intermittently. The Gaussian baseline above tests this directly at each source spacing.
Why not just add a sparsity penalty to a standard grid inversion? This is the real threat. A group-sparse grid inversion encodes most of our assumptions, which is why it is our strongest baseline. Our expectation is that it fails in two specific ways. When two sources fall inside one cell or straddle cells, grid discretization merges or smears them. It also cannot represent release height, which changes footprint shape under stable conditions. If it matches us even at sub-cell spacings, then the contribution reduces to "group sparsity across overpasses works". We would report that and drop the off-grid machinery, because the persistence-across-overpasses insight would survive either way.
Why not just subtract bottom-up wetland and inventory estimates from the top-down total? The European synthesis (The consolidated European synthesis of CH4 and N2O emissions) shows what this leaves. Even after accounting for about 5.2 Tg/yr of natural sources, top-down methane (23.3–28.8 Tg/yr) remains above the 19–21 Tg/yr bottom-up estimates. Subtraction locates nothing, and errors in the subtracted models flow straight into the residual. Our smooth background term instead estimates the diffuse component jointly from the data, and the synthetic tests score how well it is separated.
Why not just run a CNN plume detector like MARS-S2L? It is restricted to roughly 500 kg/h and above in high-resolution imagery, and it outputs detections, not rates for overlapping sources. We use it instead as ground truth.
Proposed Experiments
The first experiment is a synthetic observing-system study in the style of CERES. We simulate basins of 50–400 sources with power-law emission rates and on/off intermittency. Transport comes from HYSPLIT driven by real ERA5 meteorology for the Permian, and the simulated observations use TROPOMI-like sampling, pixel size, cloud gaps and noise. We sweep source spacing, emission rate, number of usable overpasses and noise level.
Metrics are as follows:
- Source recovery, scored as F1 for matches within 1 km at matched false discoveries.
- Location error.
- Rate error, on a log scale.
- Basin-total error.
- Recovery of the on/off pattern.
Two results are decisive. The first is that recovery for sources closer than the pixel size should rise with the spread of wind directions across overpasses, more than with the raw number of overpasses. This is the mechanism claim, and it fails if recovery depends only on count. The second is that we should beat group-sparse grid inversion specifically below about 3 km spacing; at wider spacings the two should tie.
The second experiment tests the method on real data with independent truth. We decompose TROPOMI-only observations over MARS-S2L sites and basins and ask whether recovered sources coincide with the 4,230 verified high-resolution plumes, and with their observation dates where the satellites' revisits overlap. Neither the per-overpass point-source baseline nor the Gaussian inversion has a reported score on this matching task in the file. For reference, MARS-S2L itself reports a mean average precision of 0.67 for detection in its own imagery, a number we do not compare against directly. A strong positive result would recover persistent sources below 25 t/h whose locations are later confirmed by Sentinel-2 plumes. A null result, in which TROPOMI-derived sources never coincide with high-resolution detections beyond chance, would show that real retrieval and transport errors break the approach. That outcome would also tell us whether the per-overpass wind correction is enough.
References
- Global assessment of oil and gas methane ultra-emitters
- Tutorial on MUedit: An open-source software for identifying and analysing the discharge timing of motor units from electromyographic signals
- A Biophysical-Model-Informed Source Separation Framework For EMG Decomposition
- Definitions and methods to estimate regional land carbon fluxes for the second phase of the REgional Carbon Cycle Assessment and Processes Project (RECCAP-2)
- Mesoscale inversion: first results from the CERES campaign with synthetic data
- AI for operational methane emitter monitoring from space
- The consolidated European synthesis of CH 4 and N 2 O emissions for the European Union and United Kingdom: 1990–2017