Daily Gap-Filling of TROPOMI XCH₄ at 5 km Resolution over China

Daily, high-resolution XCH₄ products are essential for verifying emission inventories and detecting transient anomalies, yet TROPOMI retrievals over China suffer from pervasive gaps due to cloud cover, albedo variability, and retrieval quality. We present a daily, gap-free XCH₄ dataset at 5 km resolution over China (2018–2025), reconstructed from TROPOMI using XGBoost augmented with Random Fourier Features (RFF). The RFF projection eliminates the rectilinear mosaic artifacts intrinsic to axis-aligned tree splits, yielding spatially continuous fields suitable for downstream applications. The model integrates ERA5 reanalysis, CAMS methane analysis, MODIS NDVI, and a secular trend term within a feature space spanning chemical, meteorological, and biospheric processes.Variogram-informed spatial-block cross-validation yields R² = 0.860 and RMSE = 13.96 ppb, whereas conventional random cross-validation inflates explained variance by approximately 44%, exposing a pervasive overestimation of predictive accuracy. Independent validation against two TCCON stations entirely excluded from the training pipeline achieves R² = 0.835 and RMSE = 12.57 ppb (N = 1,245), expanding daily matchup availability 2.4-fold and reducing systematic bias from +4.33 to +2.08 ppb relative to the original quality-filtered TROPOMI retrievals. SHAP analysis confirms that temperature and soil-moisture responses dominate the learned feature hierarchy, consistent with methanogenesis biogeochemistry. The reconstructed dataset resolves short-lived anomalies invisible in the raw satellite record, including the 2022 Yangtze drought suppression and 2020 COVID-19 lockdown perturbations over Shanxi Province. The product is suitable for regional emission monitoring, anomaly detection, and as prior fields for flux inversions.每日高分辨率XCH₄产品对于验证排放清单以及探测瞬态异常至关重要，但针对中国区域的TROPOMI（对流层监测仪器，Tropospheric Monitoring Instrument）反演结果却因云覆盖、反照率变化以及反演质量问题存在大量数据缺失。本研究构建了一套2018至2025年覆盖中国区域、分辨率为5千米的无缺失每日XCH₄数据集，该数据集基于TROPOMI原始反演结果，通过结合随机傅里叶特征（Random Fourier Features, RFF）的XGBoost（极端梯度提升，eXtreme Gradient Boosting）模型重构得到。随机傅里叶特征投影可消除轴对齐树分裂固有的矩形马赛克伪影，最终得到适用于下游应用的空间连续场数据。该模型融合了ERA5再分析资料、CAMS（哥白尼大气监测服务，Copernicus Atmosphere Monitoring Service）甲烷分析产品、MODIS NDVI（中分辨率成像光谱仪归一化植被指数，Moderate Resolution Imaging Spectroradiometer Normalized Difference Vegetation Index）以及长期趋势项，构建涵盖化学、气象与生物圈过程的特征空间。基于变异函数指导的空间块交叉验证得到的决定系数R²为0.860，均方根误差RMSE为13.96 ppb（十亿分之一）；而传统随机交叉验证会将解释方差高估约44%，这揭示了预测精度普遍被过度估计的问题。针对完全未纳入训练流程的两个TCCON（总碳柱观测网络，Total Carbon Column Observing Network）站点开展独立验证，结果显示决定系数R²为0.835，均方根误差RMSE为12.57 ppb（样本量N=1245）；相较于原始经过质量过滤的TROPOMI反演结果，该数据集将每日匹配数据可用性提升了2.4倍，并将系统偏差从+4.33 ppb降至+2.08 ppb。SHAP（夏普利可加解释，SHapley Additive exPlanations）分析证实，温度与土壤湿度响应在习得的特征层级中占据主导地位，这与产甲烷作用的生物地球化学规律一致。重构后的数据集可捕捉到原始卫星观测中无法识别的瞬态异常，包括2022年长江流域干旱抑制效应以及2020年新冠疫情封锁期间山西省的甲烷扰动信号。该数据集产品适用于区域排放监测、异常探测以及作为通量反演的先验场。