Signal-Domain Guided Deep Learning for Gap-Filling of XCO and XCH₄: A Masked Spatio-Temporal Fusion of TROPOMI and GEOS-Chem (2019–2023)

Overview This dataset provides high-resolution, gap-free global and regional atmospheric products of Column-averaged dry air mole fractions of Carbon Monoxide (XCO) and Methane (XCH₄) from 2019 to 2023. The data is generated using a novel signal-domain fusion approach that integrates TROPOMI satellite observations with GEOS-Chem chemical transport model simulations. By combining 3D Discrete Cosine Transform (DCT), Singular Value Decomposition (SVD), and a lightweight residual U-Net deep learning model, this dataset overcomes the common issue of missing values in satellite data due to cloud cover and retrieval limitations while maintaining high spatial accuracy. Data Content The repository includes two primary scales of data: Global Coverage: Daily XCO and XCH₄ at 0.25° × 0.25° resolution. Regional Coverage (China): High-resolution daily XCO and XCH₄ at 0.05° × 0.05° resolution. Key Features Continuity: 100% spatiotemporal coverage (gap-free) for the period 2019–2023. High Accuracy: Validated against TROPOMI and independent observations, achieving $R^2 = 0.92$ for XCO and $R^2 = 0.85$ for XCH₄. Enhanced Precision: Effectively captures localized events (e.g., 2022 Chongqing wildfires) and refines agricultural emission patterns in rice-growing regions compared to raw satellite data. Physical-ML Fusion: Leverages both atmospheric physics (GEOS-Chem) and deep learning (U-Net) with meteorological drivers. File Structure & Usage Format: The data is provided in .nc (NetCDF4) Variables: XCO (ppb), XCH4 (ppb), day, lat, lon. Potential Applications: This dataset is suitable for climate mitigation strategy analysis, regional emission assessments, trend analysis of greenhouse gases, and as input for atmospheric inversion models. Methodology The fusion process involves: Initial Reconstruction: Signal-domain integration using 3D DCT and SVD. Deep Learning Refinement: A residual U-Net learns the residual field between initial reconstructions and ground truths, guided by a masked loss function and meteorological inputs.