Global warming is a major challenge today, and increasing CO₂ is the main driver. Among other things, remote sensing observations from carbon satellites provide important data for XCO₂ measurements. On this basis the new generation of satellites is planned to use the strategy of wide orbit to detect high concentration areas first, and then narrow orbit for detailed study. The initial detection of high concentration areas is followed by detailed further observations. For the wide orbit low-resolution observation, the all-physics inversion algorithm has high accuracy but large computation. Based on this, this study proposes a deep learning approach and tests the method with hyperspectral observations from two satellites, OCO-2 and TanSat, which include a cloud detection method based on surface barometric pressure inversion and XCO2 inversion. The training data are obtained from surface barometric pressure data from the European Centre for Medium-Range Weather Forecasts (ECMWF) and atmospheric XCO₂ data from the Copernicus Atmospheric Monitoring Service (CAMS), and only the sky-bottom observations are selected for training and evaluation. The main input variables include O₂-band, weak CO₂-band, and strong CO₂-band data, and 30% of the data are randomly selected for training, while the remaining data are used for model validation. The results show that for OCO-2 satellite, the deviation of the method is only 0.01hPa for surface pressure inversion and almost no deviation for XCO₂ inversion, and for TanSat satellite, the deviation of the method is only 0.08hPa for surface pressure inversion and 0.10ppm for XCO₂ inversion, and it has certain advantages over the traditional all-physics algorithms in the aspects of cloud detection and high-level XCO₂ inversion. Compared with the traditional all-physics algorithm, it has certain advantages in cloud detection and high-level XCO In addition, the validation using TCCON site data further demonstrates the feasibility and potential application of the deep learning method in XCO₂ inversion. The work in this study will provide technical support for the upcoming launch of the new Carbon Star and promote the development of deep learning-based GHG remote sensing inversion techniques.
 

TanSat-2, XCO₂ retrieval, neural network, DenseNet, cloud screening, hotspot detection