The manuscript "GHGPSE-Net: A method towards spaceborne automated extraction of greenhouse-gas point sources using point-object-detection deep neural network" presents a novel deep learning framework and a large dataset for detecting and quantifying greenhouse gas (GHG) point sources from satellite imagery. The authors are the first to introduce the point object detection approach in this domain, which is very valuable and insightful for the GHG point source monitoring community, as it has the potential to integrate and simplify the processing complexity largely. The authors also demonstrate the feasibility of the model by evaluating it on two datasets, including authentic satellite observations. Though I anticipate more evaluations may be required on the upcoming moderate-resolution carbon monitoring satellites (e.g., CO2M and TanSat-2) to fully explore the potential. This work marks an important step towards automated GHG point source monitoring and has the potential to make a significant contribution to the GHG remote sensing community.

Minor suggestions:

(1-1) The dataset construction process, including WRF-GHG simulation, XCO2 construction and data augmentation, involves multiple scenarios, especially it seems that the model is trained on the synthetic dataset and evaluated using independent datasets. It may be better clarified using a diagram.

(1-2) The authors summarized GHGPSE-Net in Figure 3. However, the overall methodology, including simulation, simulation evaluation, training dataset preparation, and deep learning evaluation, is quite complex and somewhat difficult to follow. The authors may consider summarizing the entire methodology in Section 2.

(1-3) In some related GHG plume detection studies, deep learning models usually require wind as an input. Does GHGPSE-Net not require the 2D wind field as input?

(1-4) According to the result (e.g., Table 3), it seems the "2 km × 2 km" in L10 should be 0.5 km × 0.5km.

Technical comments:

(2-1) Typo in L59 and L137.

(2-2) It should be "mean squared error" instead of "mean square error" in L10, L208, and L223.

This work applies deep learning to satellite snapshots of CO2, performing point source detection, localisation and quantification. As the authors identify, most work in this area focuses on plume identification or source location, without performing quantification directly. The authors present an innovative solution, where a CNN is trained to identify sources as hotspots, which then get detected and quantified using a statistical Gaussian Kernel fitting method. The authors also introduce a greenhouse-specific augmentation method, varying the linear coefficients for each source, that presents a valuable contribution to the area of GHG+deep learning. The authors demonstrate their model on a wide range of observations and instruments, using also independent validation tests for verification. I believe that the work presents a significant contribution to the ML-remote sensing field, with some further clarifications on the method.

Clarifications required:

• The datasets used should be summarised more cohesively, as at the moment it is hard to follow 1) what regions, areas and time periods are covered for each dataset, 2) the number of samples in each dataset 3) whether the test datasets were augmented. The number of samples added in the augmentation needs to be specified.
• The architecture of the model needs to be outlined more clearly, including how it was tuned, the learning rate, loss functions etc.
• It is not clear to me if the resolution of the model was kept constant across the different instruments, or adapted to each instrument.
• Is it possible that the significantly lower skill on the SMARTCARB dataset is due to a different transport model?
• Would other inputs to the CNN, like the wind direction or the location of known sources, improve predictive skill?
• The Gaussian kernel fitting is demonstrated to improve when the kernel size is the same as the instrument pixel size, for the one instrument tested. Do you expect performance to improve for other instruments when the same process is applied? It is not clear what kernel size you applied for evaluation for each instrument.

Typos and minor corrections:

• Figure 5 shows, from the text, two days with large absolute errors. Clarify in the caption/figure labelling that these two days are not representative of the usual error distribution.
• Line 39 (“and The new generation…”)
• Line 151: areas (not area)
• L342
• Supplement S3.1, title: Definition (not defination)