The authors present a comprehensive study on the estimation of methane emission rates from livestock in sub-Saharan Africa using a combination of drone, flux tower, and satellite data. They developed a Bayesian inference method to assimilate these observations into an atmospheric dispersion model and compared the methane emission rates of various livestock species, including cattle, goats, sheep, and camels, from the Bayesian inference method with estimates from a mass balance approach and IPCC Tier 1 and 2 approaches. The Bayesian inference method was found to be more robust at quantifying emissions from weaker sources than a mass balance approach. The results indicate that the Bayesian inference method is effective in quantifying methane emissions from different livestock species, with promising implications for improving emission inventories, especially for less studied livestock species such as camels.

In general, the paper is well prepared and well-reasoned. However, there are areas where the manuscript could be improved to enhance clarity and impact. The authors have studied the topic from many angles, and thus the manuscript would benefit from better structure and reorganization of some topics. 

In particular, I wondered about the value of using the PRISMA observations. It seemed to me that the development of the Bayesian inference approach was the main thing in this paper. The use of satellite observations felt a bit like an afterthought, especially since the authors could not use it as a validation of the Bayesian inference approach. Thus, I would suggest that the authors change the order of the paper so, that the sections discussing PRISMA observations would be the last, and that they would be introduced as a demonstration how it could potential be used to verify or upscale the Bayesian inference approach, similar how you state in P3, L77-80: “We apply two distinct methods: a traditional mass balance method and an innovative Bayesian inference approach that uses a sequential Monte Carlo method to invert an atmospheric diffusion model. To complement this analysis, we assess the capability of hyperspectral satellite data to pinpoint the location of CH4 sources, specifically ruminant herds, by identifying spectral anomalies at the landscape level.” 

I was also puzzled by the "4.1 Bayesian lessons learned" section. It was under the conclusion but introduced topics that were not properly discussed in the paper. It seems that the authors did a lot of testing with the models before settling on the setup explained in the manuscript. Such testing is desirable, and I appreciate that they report what worked and what did not. However, I think it should be discussed before the conclusions, or at least there could be a mention somewhere at the beginning that the lessons learned will be discussed at the end of the manuscript, for example similar to Saunois et al, 2024 (https://doi.org/10.5194/essd-2024-115). I would keep the conclusion section itself rather short and focused.

I would also consider separating the Results and Discussion sections into two different ones. 

Specific comments 

P1, L17-18:  

“improvement of climate models”: I did not find a proper reasoning for how the methods developed in the paper would improve climate models. From my point of view, the major improvements would be within inventories and verifying other methods which estimate methane emissions. Please consider removing this or add reasoning for this. 

P2, L27: 

“a mass balance method with drone observations”: You mention later that this is an established method to estimate emissions using drone observation. You could mention it already here to highlight that the Bayesian approach is the novel aspect here. 

P2, L29: 

“Global mean atmospheric CH4 concentrations surpassed 1.90 ppm in 2022”: You could update this value, especially since the study takes place in 2024. 

P3, L65: 

“On much larger spatial scales”: Much larger scales than herds and farms? Please, specify. 

P4, section “2.1 Satellite observations for source detection”: 

I would wish to have a better motivation why you are using observations from PRISMA and not from other satellites. Has it something to do with the spatial resolution and revisit time? Did you check if other satellites had observations over the study areas during the study period? Please, add at least a short description of the PRISMA satellite (spatial resolution and revisit time etc.). 

P4, L120: 

What is a “data cube”? 

P5, L139: 

You could mention already here why you have “one flight for each emission estimation method” and why the same flights cannot be used for both methods, i.e., they need different flight paths. 

P7, L184: 

What is “model inversion”? 

P7, L185: 

“Unlike optimization”: the word “optimization” is also used with Bayesian statistics and can include minimizing a cost function (see for example the references at the end of the chapter). 

P9, L233: 

“the threshold of 1.8 ppm”: This sounds a bit low compared to the global average (above 1.91 ppm in 2024, https://gml.noaa.gov/ccgg/trends_ch4/). How did you settle on this value? You could also discuss more how the chosen background value affected the emission rate estimates. 

P12, L333: 

“with a height of 10 m”: How did you settle on this height? 

P14, section “3.1 Source detection through satellite observations” 

As I mentioned before, I think this could be the last section. In addition, to make it more relevant, you could think of adding a simple correlation calculation, i.e.  calculate how much each PRISMA pixel should have methane emissions based on the paper (emission rate x number of animals) and then see if there is any correlation between the SR anomalies. At least you could speculate more on how these satellite observations could be used to upscale the results from this paper. 

P16, L432-436: 

“Specifically, the wind direction determines the plume’s orientation, while wind speed and diffusivity influence the plume’s shape, and the emission rate determines how elevated the plume’s concentration level is above the background. We consider these four parameters - wind direction ϕ, wind speed V, diffusivity D, and emission rate q – as unknowns to be inferred.” 

This could be mentioned already in the methods. 

P16, L443: 

“we frequently observe different patterns”: Could you specify, which patterns? 

P16, L444: 

“Except for camels, the herds consist of approximately 100 to 200 animals”: How many animals did the camel herds consist of? 

P18, L484-486: 

“Observation cases (b) and (c) present an interesting topic for future study: Is it more valuable to have an anemometer on the drone to capture local and temporal wind variations, or to place an anemometer close to the source at a fixed location and 

use MOST to obtain diffusivity observations?”: Could you speculate or form a hypothesis which could be better? 

P18, L492: 

Here, you compare the values to IPCC Tier 1 values. You didn’t introduce them in the materials section, but you could do it briefly when also introducing the Tier 2 method. 

P23, L617-619: 

Were the differences between before and after grazing statistically unsignificant? 

P23, L623-625: 

“With the exception of the cows and lactating ewes drone flights, the Bayesian inference estimates pre-grazing are lower than the IPCC Tier 2 value, and the post-grazing estimates are higher than the IPCC Tier 2 value, or the uncertainty ranges of the different methods overlap.”: Are the IPCC Tier 2 values some kind of averages, i.e. they can be used to calculate annual emissions? If so, shouldn’t you then compare some kind of average of the before and after grazing emission rates with the IPCC values? 

P23, L633: 

“In comparison to the Bayesian inference method, the mass balance approach is more straightforward to implement.”: Why is that? Do you mean, for example, that the math behind it is simpler/requires less assumptions? 

P26, L716-718: 

“Future applications of the Bayesian framework for source term estimation could 

extend to diverse natural and anthropogenic sources, such as CH4 emissions from wetlands, hotspots in thawing permafrost, landfills, and wastewater disposal sites.”: This paper focused on “point-like” sources but wetlands and landfills are “sparser” and have emissions from a larger area. How well does the method introduced here suit such cases? Are there major issues that need to be addressed before Bayesian interference can be used over such areas? 

Technical corrections 

P19, L525: 

Should it be “Figure 5” and not 2? 

P19, L536:

“negative mission rate” -> “negative emission rate” 

P23, L625 and L628: 

“IPPC” -> “IPCC”