the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
A High-Precision Satellite XCH4 Inversion Method Using CBAM-ResNet18
Abstract. Amid global climate change, rising atmospheric methane (CH4) concentrations significantly influence the climate system, contributing to temperature increases and atmospheric chemistry changes. Accurate monitoring of these concentrations is essential to support global methane emission reduction goals, such as those outlined in the Global Methane Pledge targeting a 30 % reduction by 2030. Satellite remote sensing, offering high precision and extensive spatial coverage, has become a critical tool for measuring large-scale atmospheric methane concentrations. However, traditional physical inversion models face challenges, including high computational complexity, low processing efficiency, and inadequate incorporation of spatial distribution information, limiting their effectiveness. To address these shortcomings, this study proposes a high-precision XCH4 inversion method that integrates the Convolutional Block Attention Module (CBAM) with the ResNet18 neural network (CBAM-ResNet18). By leveraging shortwave infrared spectral data from the Sentinel-5P satellite and the CAMS reanalysis dataset, this approach achieves rapid and accurate XCH4 inversion. Experimental results demonstrate that the method outperforms both conventional physical models and existing mainstream techniques in terms of inversion accuracy and computational efficiency. It achieves an error of less than 2 %, meeting the stringent precision requirements for XCH4 in atmospheric remote sensing and providing a robust tool for methane monitoring.
- Preprint
(1016 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 01 Nov 2025)
-
RC1: 'Comment on egusphere-2025-3885', Anonymous Referee #1, 22 Sep 2025
reply
-
AC1: 'Reply on RC1', Yong Wan, 25 Sep 2025
reply
Dear Reviewer,
Thank you for your careful review and constructive feedback on our manuscript. We greatly appreciate your expertise and the time you invested in providing detailed comments. Your observations highlight important aspects of our work that will help enhance the scientific rigor of the paper. Below, we address each point raised in your overall comments and major concerns. We have carefully considered your suggestions and outlined planned revisions where appropriate, providing clarifications where necessary:
1. Training Strategy:
We appreciate the reviewer raising this critical point regarding our use of interpolated coarse-resolution CAMS reanalysis data as the training target, while the model input consists of high-resolution TROPOMI spectral data. To further clarify the necessity and rationale of our approach: The primary objective of our CBAM-ResNet18 model is not to fully replace CAMS, but to develop an efficient, data-driven simulator that bridges raw TROPOMI L1B spectral observations with reliable XCH₄ estimates while incorporating the spatial context of high-resolution inputs. Several compelling reasons based on data quality, bias mitigation, and methodological advantages justify selecting CAMS reanalysis data as the training target:
(1) CAMS integrates a comprehensive collection of global observations (including satellite, ground-based, and in-situ data) with advanced chemical transport modeling, providing a more robust and bias-corrected XCH₄ representation than the standalone TROPOMI L2 product. Existing studies indicate that TROPOMI L2 XCH₄ may exhibit regional biases due to errors in prior conditions within the inversion algorithm, such as surface reflectance priors and aerosol corrections. By training on CAMS, our model learns a more accurate target, focusing on spectral features relevant to real atmospheric variability. Our experiments comparing results against TCCON ground data also validate that this approach yields more accurate results than TROPOMI L2 XCH4.
(2) Although CAMS has coarse resolution, our input preprocessing explicitly addresses this by cropping TROPOMI L1B data into 3×3 spatial blocks. This enables the model to capture subgrid heterogeneity (e.g., local emission hotspots) smoothed out by CAMS. The CBAM module further amplifies this by selectively focusing on spatial and spectral features within the high-resolution spectrum, allowing the model to “upsample” fine details onto the CAMS grid cells rather than simply reproducing their smoothing.
(3) While our data filtering step utilizes the TROPOMI L2 product XCH4, this is solely for a convenient method to screen high-quality TROPOMI spectral samples during model training. This step is not required for the model's practical application.
We acknowledge that our methodology may warrant more detailed elaboration, but we do not plan significant modifications to the core approach. Our objective is to surpass TROPOMI's L2 product XCH4 in both inversion efficiency and accuracy. Using TROPOMI's L2 product XCH4 as the training target would substantially diminish the research's necessity, and the inversion accuracy would likely not exceed that of the product. Nevertheless, to address your concerns, we will expand our explanation in Section 2, detailing why CAMS was chosen as the training target instead of TROPOMI's L2 product XCH4.
2. Model Performance:
We appreciate the reviewer's emphasis on the need for explicit data coverage details, a larger sample size, and more comprehensive metrics. We also note your observation of “flat” scatter plots in Figures 6 and 7, which may suggest low sensitivity to XCH4 variability. Upon re-examination, this result stems from the narrow dynamic range of XCH4 in our selected validation subset (1840–1850 ppb globally during June 2020), where natural variability was moderate (approximately 10 ppb), leading to y-axis compression. However, RMSE/MAE values (e.g., 9.96 ppb relative to CAMS, 5.68–17.6 ppb relative to TCCON) indicate that relative accuracy within this range is not low, and spatial visualizations of the model (Figure 5) confirm its ability to resolve subtle gradients.
To strengthen the persuasiveness of this section's experiments and enhance the model's generalization capability, we plan to implement the following modifications:
(1) To mitigate the small sample size, we will expand the dataset by incorporating more TROPOMI orbits and extending coverage to a global scale, followed by retraining the model.
(2) To comprehensively evaluate the accuracy of experimental results, we will add correlation coefficients (r) and coefficient of determination (R²) to all scatter plots (Figures 6 and 7).
3. Comparison with Other Methods:
We appreciate the reviewer's feedback and acknowledge that our previous description of the “optimal estimation” method was indeed overly broad. In the revised manuscript, we have provided further details on the method's configuration: Specifically, we employ the SCIATRAN radiative transfer model as the forward model, with CH₄ prior profiles derived from spatiotemporally interpolated CH₄ profiles from the CAMS reanalysis product; Additionally, other key parameters include surface albedo and temperature/humidity profiles provided by the ERA5 reanalysis dataset. All other parameters are configured according to SCIATRAN's default settings. We believe this supplementary information enhances the method's transparency and reproducibility.
We understand the reviewer's concern regarding the citation source for the “improved spatial inversion method.” This method was originally derived from a master's thesis within our team. While the thesis underwent rigorous academic defense and peer review procedures, as you noted, theses may be less accessible and subject to less stringent review compared to formal journal articles. Therefore, in the revised manuscript, we plan to add citations to relevant journal articles on “spatial inversion methods” and provide further clarification on the specific principles of the “improved spatial inversion method.” The forward model, prior data, and other parameters of this method remain consistent with the “optimal estimation” approach.
4. Lack of Discussion Section:
We fully agree with the reviewer's feedback that a dedicated discussion section is essential for contextualizing our contributions, implications, and limitations. While we combined the discussion with experimental analysis in a subsection, the discussion was indeed insufficient. In the revised manuscript, we plan to insert a new subsection before the conclusions to discuss experimental results. This section will include feature importance analysis based on SHAP values to identify which input features most significantly impact model predictions, thereby enhancing model interpretability.Citation: https://doi.org/10.5194/egusphere-2025-3885-AC1
-
AC1: 'Reply on RC1', Yong Wan, 25 Sep 2025
reply
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
1,600 | 23 | 12 | 1,635 | 25 | 26 |
- HTML: 1,600
- PDF: 23
- XML: 12
- Total: 1,635
- BibTeX: 25
- EndNote: 26
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1
[General Comments]
This manuscript introduces a deep learning approach, CBAM-ResNet18, for retrieving column-averaged dry-air mole fraction of methane (XCH4) from TROPOMI satellite spectra. The topic is of considerable interest and falls well within the scope of Atmospheric Measurement Techniques. The goal of developing more efficient retrieval algorithms is an important one for the atmospheric and climate science community.
However, the study has several major issues that undermine its scientific quality. These include the model training strategy, unconvincing validation performance, insufficient detail in comparison between methods, and a lack of discussion on the model's limitations. For these reasons, the work does not meet the scientific standards required for publication in the AMT journal.
Major Concerns: