The authors reconstructed two-decade global daily high-resolution XCO2 data based on a hybrid Transformer–BiLSTM model. However, the topic is usual, and the method should be more innovative in this study, which requires large improvements. Specific comments are as follows:

Major comments:

1. There are many studies providing global daily high-resolution XCO2 for a long term, such as DOI: 10.1016/j.eiar.2025.108146 with better spatial resolution (1 km) than yours (2003−2023, R² = 0.988, RMSE = 1.10 ppm against TCCON) [1]. What are your strengths compared to them? What is it that your dataset can present but others cannot? Please state in detail.
2. The BiLSTM and attention model have been applied to estimate global XCO2, such as DOI: 10.5194/essd-17-5355-2025. It seems that your model structure presenting similarity to this paper [2]. Please justify your innovation.

Minor comments:

1. Line 180: Please add some ablation experiments of your model.
2. Figure 6: A station presents very low R2 value (< 0.1). Please add some discussions.
3. Figure 8: The selected scenarios are only located on one or two pixels. I think the high values also may be some noises.
4. Figure 14: I notice the mean SHAP value of NO2 is the smallest. How to interpret this result?
5. Table 2: Some important works are missing, such as [1] and [3]-[5].

References:

[1] Wang, J. (2026). Global daily 1 km gapless XCO₂ (2003− 2023) derived from multi-satellite observations and a spatiotemporal deep learning framework. Environmental Impact Assessment Review, 117, 108146.

[2] Wang, Z., Zhang, C., Shi, K., Shangguan, Y., Hu, B., Chen, X., ... & Zhang, Q. (2025). A full-coverage satellite-based global atmospheric CO 2 dataset at 0.05° resolution from 2015 to 2021 for exploring global carbon dynamics. Earth System Science Data, 17(10), 5355-5375.

[3] Li, J., Zhang, Z., Li, T., Yuan, Q., & Zhang, L. (2026). Global daily seamless XCO2 Mapping (2016–2020): Spatio-temporal trends and variations during wildfire events. International Journal of Applied Earth Observation and Geoinformation, 146, 105092.

[4] Hwang, S., Choi, H., Kang, Y., & Im, J. (2026). Reconstructing long-term (2003–2019) global high-resolution XCO2: bridging observational gaps with machine learning. GIScience & Remote Sensing, 63(1), 2627042.

[5] Yu, Y., Tian, W., Zhang, L., Yu, T., Wu, Y., & Cheng, T. (2026). MCF-XCO2: A cross-mission consistency and fusion framework for integrating multi-satellite XCO2 observations. Atmospheric Research, 108747.

Qu et al. presents a timely and potentially valuable high-resolution daily XCO2 dataset by combining multi-mission satellite retrievals with a hybrid Transformer–BiLSTM framework, followed by a bias-correction step. The topic is well suited to ESSD, and the dataset has clear potential applications in carbon-cycle analysis, emission monitoring, and climate studies. Overall, the manuscript is well written, and I have several suggestions to further improve the study.

Specific comments:

The spatial scope of the dataset should be described more consistently. The title and several sections of the manuscript give the impression of a fully global XCO2 product, whereas the abstract states that the dataset characterizes XCO2 “over global land surfaces.” This scope should be made consistent across the title, abstract, main text, and dataset description.

The description of the Bi-LSTM architecture needs clarification. The statement that the model “consists of 64 hidden layers with a dimension of 128” is difficult to interpret and seems unlikely in its current form. It would be helpful to clarify whether this refers to hidden units, hidden states, or stacked layers.

To enhance readability, Figure 5 should include a corresponding legend.

Several language and wording issues should be refined. For example, “till 2012” should be revised to “until 2012,” and “Electric power plant in Algerian” should be corrected to “in Algeria.” A careful language edit would improve overall readability.

The ENSO discussion contains a minor numerical inconsistency. A reported correlation of R = 0.55 corresponds to approximately 30% explained variance rather than ~34%. Please check the correlation coefficient or revise the corresponding statement accordingly.

The performance metrics reported in the abstract and main text would benefit from clearer interpretation. Different statistics are presented for independent TCCON validation, bias-corrected results, and various cross-validation strategies, but the abstract may give the impression that all values correspond to a single evaluation setting. It would be helpful to explicitly indicate which metric corresponds to which validation framework.

In the caption for Figure 7, the years are listed as “May 30, 2009, 2013, 2020.” For grammatical correctness, this should be revised to “May 30, 2009, 2013, and 2020.”

The seasonal averages should be presented in a consistent format, for example (MAM; average = 395.36 ± 2.20 ppm).

The terminology used for the reconstructed product should be standardized. Terms such as “seamless,” “gap-free,” and “gapless” are used interchangeably; selecting one consistent term would improve clarity.

Some minor grammatical issues should also be corrected, for example “Sample-based CV show” should be revised to “Sample-based CV shows.”

The model name should be used consistently throughout the manuscript. In most sections it is referred to as “Transformer–BiLSTM,” whereas in the XAI section it appears as “4D-STransformer-BiLSTM,” which may confuse readers.

It would also be helpful to clarify whether coarser-resolution products were regridded prior to comparison in Figure 7.

The interpretation of localized XCO2 hotspots should be phrased more cautiously. Some descriptions appear overly definitive for a column-integrated quantity that can also be influenced by atmospheric transport and meteorological conditions. Expressions such as “consistent with localized enhancements associated with…” would be more appropriate unless supported by stronger independent evidence.

The formatting of several references should be carefully checked, as there are minor issues such as missing spaces or compressed publisher/year formatting.

In Table 2, it would be beneficial to include comparisons with the most recent relevant studies to better position the dataset within the current literature.

Finally, the names of datasets and networks should be presented consistently. For example, “ObsPack” and “Obspack” appear in different places and should be unified.