A Frequency-Domain Variable Dependency-Guided Transformer for Multivariate Greenhouse Gas Forecasting from TCCON Remote Sensing Observations

Accurate monitoring and forecasting of atmospheric greenhouse gas (GHG) concentrations is critical for climate policymaking and global carbon management. The Total Carbon Column Observing Network (TCCON) is a standardized ground-based remote-sensing network that provides accurate column concentrations of XCO2, XCH4, XCO, XN2O, and XH2O. However, forecasting TCCON GHG concentrations remains challenging because of inter-gas scale disparities, dynamic cross-gas dependencies, and multi-scale temporal variations. To address these challenges, we developed an innovative forecasting framework, the Reversible Decomposition-Patch Variable Transformer (ReDP-VT), for accurate GHG forecasting. ReDP-VT combines reversible instance normalization (RevIN) to reduce scale disparities, decomposition-patch embedding (DPE) to extract long-term and local temporal features, and frequency-domain variable relationship modeling (VRM) to encode inter-gas dependencies that guide Transformer self-attention. We conducted comparative evaluations against seven baselines across four TCCON sites. ReDP-VT outperformed all baselines at Xianghe and secured 21 best and 6 second-best results among the 27 cross-site generalization tasks at Edwards, Lauder, and Park Falls. Furthermore, gas-specific analyses showed that XCO2 was forecast with the highest accuracy, whereas XCO was the most challenging gas to model. These findings underscore the efficacy of ReDP-VT for GHG concentration forecasting, with potential applications in high-precision atmospheric monitoring and data-driven climate-policy formulation.