GENERAL:

I find this paper very interesting because it manages to shed considerable additional light on many of the big questions about the discrepancy between bottom-up and top-down estimates of methane emissions from North American sources. Without being an expert on inverse modelling myself (but rather bottom-up modelling), I still note that the authors make important improvements in the methodology that are additional to previous studies, i.e., using both satellite and in-situ observations, using a log-normal error function which better represents the high tail emission distributions that are typical for the oil and gas sources, and using an improved prior for wetlands, which does not overstate wetland emissions as has previously been a problem. These improvements seem to lead to results that better explain the total contribution from anthropogenic sources and their attribution to individual source sectors. The paper is also well written and easy to follow and I support publication but would like to see one major concern addressed and a few minor revisions, as listed below.

MAJOR CONCERN:

Authors are able to show that anthropogenic CH4 emissions are substantially underreported in all three countries USA, Canada, Mexico, and in particular for the US. They conclude that in particular emissions from oil production are underreported by a factor of 2. Looking at the trend 2010-2017 for the US, they conclude that CH4 emissions appear to have peaked in 2014 and thereafter slightly declined with the overall trend for the period still slightly increasing. This is in contrast to US official reporting to the UNFCCC, where emissions decline steadily over the period. The authors attribute the increases they find to oil production and landfill, while emissions from gas production are said to decline (and livestock and coal mining stay flat). Given that according to EIA, US shale gas production increased by 227% (from 165 to 540 bcm) over this period while oil production increased by a more modest 71% (and other natural gas production declined by 39% from 493 to 300 bcm), I am not convinced about the authors’ split in attribution between oil and gas sector emissions. I wonder if the inversions can really make this distinction between oil and gas sources as fields in the US are often producing both oil and gas? If authors are not able to do this split in a robust manner, then I would recommend the authors not to report oil and gas sector emissions separately, because from a policy point of view this matters a lot. If there is a risk that authors are wrong about their conclusions here and that in reality it is a strong increase in methane emissions from shale gas production that is picked up (and not oil), then you risk sending the completely wrong signal to policy-makers (i.e., “fix oil but don’t’ worry too much about gas production”, when it could be that the real problem is the shale gas). So if there is uncertainty regarding this, then report oil and gas emissions together and leave it to further research to figure out this split in more detail.

MINOR CONCERN/EDIT:

p.11 row 387: write out the acronym DOFS.

p.14 row 530: It is suggested that the downward correction for offshore operations can be referred to that methane from offshore oil platforms is piped onshore and inefficiently flared. Another possible explanation could be that when methane leaks happen at the seabed, methane oxidises to CO2 in the water column before reaching the surface and therefore emissions are considerably lower during offshore production. Could this be an explanation here?

General comments

The authors extend their previous coarse-grid global inversions to a fine-grid regional scale. They optimize methane emissions and 2010-2017 emission trends in North America by in situ (GLOBALVIEWplus CH₄ ObsPack) and satellite (GOSAT) observations, through analytical inversions using log-normal error forms. They point out large emission underestimates in the oil sector by a factor of 2, and a peak of CONUS anthropogenic emissions in 2014. The paper is well written. The methods are clearly described, and the results are well discussed. I support publication, but with a major concern and some minor suggestions.

My major concern is that this study lacks independent evaluation. The authors compare the posterior simulation and prior simulation against the observations used for the inversions, and the improvements against GOSAT are weak. I am curious about the evaluation against the independent dataset, such as TCCON or other local in-situ measurements.

Specific comments and technical corrections

1. Row 199: What are the treatments for the initial conditions of the global simulations?
2. Row 254: Are the Jacobian matrix for the boundary conditions constructed in the same way as the grid-level emissions?
3. Row 297: The error standard deviations for boundary conditions are 10 ppb in the base inversion and 5 ppb in the sensitivity inversions. These are much smaller than the error standard deviations for emissions. How sensitive are the results if applying a larger error standard deviation for boundary conditions?
4. Row 317: The observation error standard deviations for in-situ data are ~2× of that for GOSAT, and the total number of observations for in-situ data is 0.4× of that for GOSAT. Readers may be curious about the results if applying two regulation parameters separately to in-situ and GOSAT data.
5. Row 417: “This may reflect the underestimation of CO2 over the Los Angeles Basin”. Typo, CH₄ rather than CO₂?

6. Row 425: “For GOSAT the improvement is less apparent from the comparison statistics, because the prior simulation already has a low mean bias MB = -0.5 ppb, and the prior RMSE is only 6.9 ppb (which decreases to 6.5 ppb). However, we see from Figure 5 a significant whitening of the noise with reduction of regional-scale biases.” This sounds like to contradict itself. It first presents that the regional mean bias of the posterior simulations is 0.6 ppb, larger than the -0.5 ppb in the prior simulations, then indicates that the posterior simulations' regional-scale bias is less than prior simulations in the map.