Joffrey et al. have developed a framework for applying CNN to estimate the power plant CO2 emissions from satellite imagery of the CO2 total column mixing ratios (XCO2). The model appears to be both resilient to common data challenges (e.g., cloud cover and limited satellite swath), and effective, with an initial validation on synthetic data from eight German power plants. Results from the application to real satellite data show a promising alignment with reported annual emissions, suggesting that the approach could be viable for broader CO₂ monitoring efforts. The authors have done a commendable job of presenting the data. Overall, this study presents a compelling and innovative approach to monitoring CO₂ emissions from power plants via satellite imagery, leveraging the power of deep learning. With some refinements in clarity and additional context around validation, this work has the potential to make a significant contribution to remote sensing and environmental monitoring fields.

 minor suggestion:

Quantification of Model Performance: Although the abstract mentions relative errors "close to 20%," it would be beneficial to briefly mention how these results compare to traditional or alternative inversion methods if applicable.

Dumont Le Brazidec et al. introduced an innovative approach to quantifying CO₂ emissions from atmospheric plumes observed in satellite imagery using Convolutional Neural Networks (CNNs). Drawing from a previous study conducted by the same team in 2023, the work represents a significant step in combining synthetic and real-world datasets to train models capable of reasonably estimating emissions, even under challenging conditions such as cloud cover or limited plume visibility. The study's topic is relevant and crucial given the challenges of extracting key information, such as local emission rates from remote satellite sensing. Integrating data augmentation, using ancillary data (e.g., NO₂ and wind fields), and applying interpretability tools (e.g., integrated gradients) demonstrate methodological novelty and potential for real-world applications. In my opinion, while the study presents promising results and is well-aligned with the scope of Geoscientific Model Development (GMD), certain aspects are yet to be improved for the manuscript to qualify for publication in the GMD. The revision suggestions, although multiple, qualify as "minor." Please refer to the comments below for more details. 

Technical Comments:

I) On figures and their captions, in General: Most figures provide insufficient information on the units of the depicted physical property. The authors should clearly state the units, as they have done for Figures 1 and 3. Figures 9 and 10 are the most severe examples of this issue, where neither mixing ratios nor emission rates come with the proper units. Furthermore, multiple instances exist where the figure's details have not been appropriately explained in the caption. Note that since many readers will engage with the publication by checking out the figures, the captions need to be clear enough, even when clarifying explanations are mentioned in the text. For instance, in Figures 3, 9, and 10, the year to which the XCO2 map belongs should be noted. Also, in Figure 8, the outlier designation criterion should be clearly explained in the caption. In other words, the authors should be explicit about the percentiles to which the box plot whiskers correspond. 

II) On subscript notation: The subscripts are managed inconsistently throughout the manuscript. For example, XCO₂ and NO₂ are also written as XCO2 and NO2. 

III) Citation system: The citation method needs proofreading. There are instances of putting the publication year within parentheses, while the parentheses have been omitted elsewhere. 

IV) On the relation between Section 3.1.1 and Figure 4: This is another example where the reader should bounce back and forth between the text and the figure to understand the content properly. Although the layers in Figure 4 are for demonstration, it's still recommended to put color legends for the sake of completion. The noise layer lacks details, and the reader would not recognize it as a Gaussian noise with a given standard deviation unless reading the text. The text in Section 3.1.1 mentions augmenting the main data with either the beta or the uniform distribution; however, the scheme only shows the beta. The purple box in Figure 4 seems misleading. Rather than sequential execution, the relation between 4 and 5 is more characterized by the iterative nature of parameter adjustment during the optimization. A feedback arrow from 5 to 4 can better show the latter. This approach will reinforce the discussions in 3.1.5.

V) Discussion around Figures 9 and 10, tortuous and difficult to follow: The panels of Figure 9 represent three conditions associated with panels (a)-(d), (e)-(f), and (g)-(h), discussed in the surrounding text. These panels are samples out of a bigger dataset. Why these eight and not the others? Have you inspected all the snapshots, discovered these three trends, and chosen to present these eight panels as "supporting examples"? If yes, please include some explanation in the text. Moreover, If you have examined other snapshots, aren't other significant plume trace-emission estimate trends to be presented? If you have only examined these eight snapshots, please advise why a broader inspection is not warranted if not required. I posted the comments above because I find it confusing to follow the overarching logical flow here, as I have difficulty understanding the natural order of evidence and conclusion. The discussion becomes even more dense and complex in lines 297 to 305. I see a complex issue being presented questionably. The logic is partially valid but could be problematic for the following reasons:

1 Dependence on Background Contrast:

◦ If the CNN interprets more pronounced plumes as higher emissions, it should perform well when the plume is distinct from the background. However:

▪ The authors argue that lower contrast in real plumes leads to underestimation.

▪ This implies that the CNN might not sufficiently generalize from simulated data to real-world conditions where plumes may naturally blend more with the background.

2 Flaw in Justification:

◦ The authors use the difference in plume-background contrast to explain underestimation. However, this reasoning partially undermines their earlier claims that CNN performs well on high-emission cases in synthetic data.

◦ The CNN was trained on synthetic images with high-emission plumes designed to be pronounced. If real-world high-emission plumes are less distinct, the model's training data might not adequately represent the variability in actual observations, leading to systematic bias.

3 Mismatch Between Training and Real Data:

◦ If the CNN was trained to interpret pronounced plumes as high emissions, it might struggle when the plume signal is weaker (as with actual data). This is consistent with their explanation but also highlights a limitation in their approach: a lack of robustness to discrepancies between simulated and real-world data.

To summarize this long comment, the mismatch between simulated and real-world data, especially in plume-background contrast, likely leads to systematic errors in CNN. While the authors' reasoning is plausible, it would benefit from a more substantial acknowledgment of these limitations and a discussion of how to address them in future work.

Specific comments:

Line 56: This sentence, "In previous studies, the models have only been tested with synthetic images without missing data," needs a citation. 

Line 82, item ii: This explanation is ambiguous and begs further clarification. • The phrase "satellite swath constraints" could refer to:

◦ The limited spatial extent of a single swath.

◦ Gaps in coverage between consecutive swaths.

◦ Temporal constraints where a given area is not revisited frequently.

•A reader unfamiliar with OCO-3 or the specific satellite mission might be left uncertain about what drives the need for a finer grid.

Table 1: Please indicate the year(s) for which the data in the third column correspond. If multiple years are involved, include information on data variability.

Line 116: Please provide exemplifying instances for "... this approach has limitations."

The squares in Figure 3: Orange and red shades look too similar here. Why not show one square with dashed sides and the other with solid sides?

Lines 158 to 160: The statement "either uniform or beta" is unclear. Please elaborate more. Is the choice between beta and uniform completely random? Figure 5 shows that the choice potentially leads to widely different coefficients, which is not a problem as you are trying to synthesize a vast artificial input database to be passed to the CNN but still begs questions about the details of the procedure.

Line 177: "... ranging from 0.75 to 2..." Is the sampling from a uniform distribution?

Table 2: For clarity, mention that the table entries are percentiles. 

Lines 237 and 238: I find this approach to presenting results awkward. Usually, when some factors do not turn out to be influencing, this condition is described in the text, and the graphs are plotted without including them in the procedure. However, one can still post the graphs, including the minor factors, in the paper's supplementary material. For your choice of manuscript configuration, I suggest showing the plots without the wind/date effect while mentioning this point in the text and the figure captions.

Lines 239 and 240: It would illustrate your quantitative criterion for calling a difference significant. A difference of < 5% seems to have been deemed significant enough to be discussed in lines 235 to 238. Figures 6 and 7 show values as low as the number below for Lippendorf and Boxberg across the mentioned scenarios of cloud coverage.

Line 250: Still confused about whether wind fields and time information are required for the algorithm to function adequately.

Table 3: I recommend removing this table as redundant because Table 1 and Figure 8 already address all the information it offers.

Lines 280 to 289: This is improperly referring to a numbered item. Please consider using Figure 9(a), Figure 9(b), and so on instead of Images (a), (b), etc.

Section 6: The manuscript would benefit from more explicitly acknowledging its limitations and possible future directions. The authors could better emphasize how their approach improves upon traditional inversion methods beyond computational efficiency. In particular, while the manuscript includes references to related work, there is insufficient discussion on how this approach compares quantitatively with other lightweight or traditional inversion techniques regarding accuracy and computational efficiency. Moreover, the limited generalizability of the CNN to real-world data due to training biases is acknowledged but not addressed in-depth. Incorporating hybrid training datasets (real and synthetic) is suggested but not explored experimentally.