Abstract:

Carbon dioxide (CO₂) is a primary greenhouse gas that plays a crucial role in global warming. Accurate information on the spatial distribution of CO₂ is imperative for assessing the efficiency of carbon emission reduction, which is important for promoting carbon peaking and neutrality. Compared to station observations, carbon satellites can obtain large-scale information on CO₂ distribution. However, the narrow width and cloud coverage associated with atmospheric CO₂ satellite remote sensing data contribute to numerous gaps, hindering the acquisition of spatially continuous atmospheric CO₂ distribution. Using the Xinjiang Uygur Autonomous Region as the study area, this study synthetically compared the performance of several methods in generating spatially continuous atmospheric XCO₂ data. Multiple linear regression (MLR) models, geographically weighted regression (GWR) models, support vector machines (SVR), random forests (RF), extreme gradient boosted trees (XGBoost), and extreme random trees (ERT) were employed to produce gap-free XCO₂ data based on the column-averaged dry-air mixing ratio of CO₂ (XCO₂) data derived from the 2019 OCO-2 satellite and auxiliary variables, such as temperature, topography, vegetation, and atmospheric NO₂ concentrations. The cross-validation results showed that the three integrated learning models, RF, XGBoost, and ERT, outperformed the SVR, GWR, and MLR models. The ERT model achieved the highest accuracy with an R² of 0.748 and a mean absolute error of 0.489×10⁻⁶. Based on the developed ERT model, spatially continuous atmospheric XCO₂ data was mapped monthly across Xinjiang in 2019. Significant spatial differences were observed in the atmospheric XCO₂ within the region. The spatial distribution pattern of XCO₂ exhibited a distinct correlation with the topographical characteristics commonly referred to as “three mountains and two basins.” Bare land exhibited the highest annual mean XCO₂, whereas woodland and grassland exhibited the lowest values. Seasonal fluctuations in XCO₂ coincided with the phenological cycles of vegetation, with peaks observed during spring and winter, and troughs during summer and autumn. The highest value was observed in April, whereas the lowest was observed in August. Across different land cover types, the seasonal variation in XCO₂ was similar to the regional patterns, with grassland and woodland experiencing the most pronounced seasonal disparities and bare land the least. This study provides a valuable framework for generating spatially and temporally continuous large-scale XCO₂ data using satellite XCO₂ products.

Key words: OCO-2, XCO₂, satellite remote sensing, integrated learning model, spatialization, spatiotemporal distribution, Xinjiang Uygur Autonomous Region

摘要：

二氧化碳（CO₂）是导致全球变暖最主要的温室气体，掌握准确的CO₂空间分布信息可以有效评估碳减排成效，对于推进碳达峰、碳中和工作具有重要意义。相比站点观测，碳卫星能够获取大尺度的CO₂分布信息，但是由于其幅宽较窄以及云覆盖的影响，大气CO₂卫星遥感数据存在大量缺值区域，不能获得空间连续的大气CO₂分布。以新疆维吾尔自治区为研究区，基于2019年OCO-2卫星大气二氧化碳柱平均干空气混合比（XCO₂）数据，结合气温、地形、植被、大气NO₂浓度等相关变量，综合对比了多元线性回归（MLR）、地理加权回归（GWR）、支持向量机（SVR）、随机森林（RF）、极端梯度提升树（XGBoost）和极端随机树（ERT）等方法在生成大气XCO₂空间连续数据中的表现。交叉验证结果表明，RF、XGBoost和ERT这3种集成学习模型精度明显优于SVR、GWR和MLR模型，其中ERT模型精度最高，决定系数R²为0.748，平均绝对误差σ_MAE为0.489×10⁻⁶。基于ERT模型生成了新疆2019年逐月大气XCO₂空间连续数据，该地区大气XCO₂空间差异和季节变化都很明显。在空间分布格局上总体呈现与地形“三山夹两盆”相似的空间分布特征，裸地XCO₂年均值最高，林地和草原的XCO₂年均值最低；季节变化特征与植被生长周期相似，春冬季浓度较高，夏秋季浓度较低，最高值出现在4月，最低值出现在8月，不同土地覆盖类型的大气XCO₂季节变化与全区域类似，其中草原和林地的季节差异最大，裸地季节差异最小。该研究为基于碳卫星遥感数据生成时空连续的大尺度XCO₂数据提供了参考和借鉴。

关键词: OCO-2, XCO₂, 卫星遥感, 集成学习模型, 空间化, 时空分布, 新疆维吾尔自治区