XCO₂ retrieval using dual-branch convolutional network with OCO-3 hyperspectral and ERA5 meteorological data

In order to address the limitations of traditional physical retrieval algorithms in accurately estimating XCO₂ under complex aerosol conditions, we propose a multi-source data fusion method integrating OCO-3 satellite L1B spectral data, ERA5 reanalysis data, and ground-based validation data from the Total Carbon Column Observing Network (TCCON). Heterogeneous feature inputs were constructed by combining shortwave infrared hyperspectral radiance characteristics with vertical atmospheric state parameters, while a spatiotemporally matched dataset was established using TCCON site XCO₂ measurements as labels. Building on this framework, we designed a Hyperspectro-Meteorological Dual-Branch Network (HM-DBNet). This architecture employs 2D convolutions to extract spectral-spatial features and integrates 1D convolutions to model vertical distributions of meteorological parameters, achieving end-to-end mapping from spectroatmospheric multimodal data to XCO₂ concentrations. Validation experiments at two Chinese TCCON sites demonstrate that our method achieves a mean XCO₂ retrieval error of less than 0.5 ppm(MAE).