Atmospheric CO₂ retrieval from satellite observations is often hindered by orbital variability, clouds, and aerosols, resulting in degraded accuracy under complex aerosol conditions. To address these challenges, this study introduced a CO₂ Concentration Feature-Merging Dual-Branch Network (CCFM-DBNet), a dual-branch feature-enhanced architecture designed for precise atmospheric CO₂ inversion. The spectral branch integrated a Global Grouped Coordinate Attention (GGCA) module into a 2D convolutional backbone to strengthen responses to key CO₂ absorption bands while suppressing aerosol and water-vapor interference. The meteorological branch employed a Multi-Scale Residual Attention (MSRA) module that combines multi-scale convolutions, residual connections, and channel attention to effectively capture vertical meteorological profiles. Feature vectors from both branches were fused along the feature dimension, enabling complementary physical information and achieving end-to-end XCO₂ prediction. Experimental results demonstrated that the CCFM-DBNet increased the coefficient of determination (R²) by 0.65% and reduced RMSE and MAE by 7.83% and 9.02%, respectively, compared with the baseline model. These improvements confirmed the method’s effectiveness in enhancing XCO₂ retrieval accuracy under complex atmospheric conditions.
CHENY F, LUBABUM M. Investigating the interrelationship among renewable energy consumption,green finance,agriculture,green technology,industry,and urbanization on CO2 Emissions: Evidence from BRICS-M-a countries[J]. Renewable Energy,2026,256:124052.
ZHANGT, ZHANGW C, YANGR Z,et al. CO2 capture and storage monitoring based on remote sensing techniques: A review[J]. Journal of Cleaner Production,2021,281:124409.
DANJOUA, BROQUETG, LIANJ H, et al. Evaluation of light atmospheric plume inversion methods using synthetic XCO₂ satellite images to compute Paris CO₂ emissions[J]. Remote Sensing of Environment, 2024, 305: 113900.
LIUW C, LIR, CAOJ,et al. Mapping high-resolution XCO2 concentrations in China from 2015 to 2020 based on spatiotemporal ensemble learning model[J]. Ecological Informatics,2024,83:102806.
[16]
TIANW J, ZHANGL L, YUT,et al. Using multisource data and time series features to construct a global terrestrial CO₂ coverage by deep learning[J]. IEEE Transactions on Geoscience and Remote Sensing,2024,62:4109814.
[17]
HUJ, SHENL, SUNG. Squeeze-and-excitation net‑works[C]//IEEE Computer Society. 2018 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Salt Lake City:IEEE,2018:7132-7141.
[18]
WOO S, PARKJ, LEEJ Y, et al. CBAM: convolutional block attention module[C]//European ComputerVision Association. 2018 Proceedings of the European Conference on Computer Vision (ECCV). Cham: Springer, 2018: 3-19.
[19]
HEK, ZHANGX, RENS, et al. Delving deep into rectifiers: surpassing human-level performance on ImageNet classification[C]//IEEE Computer Society. 2015 IEEE International Conference on Computer Vision(ICCV). Santiago: IEEE, 2015: 1026-1034.
[20]
GLOROTX, BENGIOY. Understanding the difficulty of training deep feedforward neural networks[C]//JMLR Workshop and Conference Proceedings. 2010 International Conference on Artificial Intelligence and Statistics (AISTATS). Sardinia: JMLR, 2010: 249-256.
[21]
UNDBERGS M, LEES I. A unified approach to interpreting model predictions[C]//NeurIPS Foundation. 2017 Advances in Neural Information Processing Systems (NeurIPS). Long Beach, CA: Curran Associates, 2017: 4768-4777.
[22]
ZHANGW H, WANGZ Y, LIT,et al. Random forest-based retrieval of XCO2 concentration from satellite-borne shortwave infrared hyperspectral[J]. Atmosphere,2025,16(3):238.
WEIW, BAOZ Y, SUY,et al. Research of rapid retrieving models for GOSAT XCO2 based on machine learning[J]. Computational Geosciences,2025,29(5):35.
[26]
HES C, YUANY B, WANGZ H,et al. Machine learning model-based estimation of XCO2 with high spatiotemporal resolution in China[J]. Atmosphere,2023,14(3):436.
[27]
CAIK, GUANL Y, LIS S,et al. Full-coverage estimation of CO2 concentrations in China via multisource satellite data and Deep Forest model[J]. Scientific Data,2024,11:1231.