Abstract:

Atmospheric carbon dioxide (CO₂) is a key driver of global climate change. To better understand its role in the Earth’s climate system, it is essential to precisely grasp the strategic layout for reducing pollution and carbon emissions and achieving regional peak CO₂ levels. This comprehensive strategy not only helps reduce greenhouse gas emissions but also promotes effective environmental improvement and sustainable development, thereby effectively addressing the threats and challenges posed by global climate change. Using the Random Forest (RF) model and remote sensing data from the Orbiting Carbon Observatory-2 (OCO-2) satellite, we constructed a detailed dataset of the column-averaged dry-air mole fraction of atmospheric CO₂ (x_CO2) for continental China in 2023, featuring monthly temporal resolution and spatial continuity. Due to the high spatial resolution of the satellite swath data, the sample size was excessively large, with data points overly concentrated in certain regions, leading to potential overfitting during the simulation. Consequently, we resampled the observational data using an averaging method to match the spatial resolution of the auxiliary variables, resulting in a dataset resolution of 0.1°×0.1°. The processed sample data were then input into the model for analysis, employing eight feature variables as auxiliary data: Normalized Difference Vegetation Index (I_NDVI), wind speed components (u₁₀, v₁₀), temperature (t), total precipitation (P_t), nighttime light index (I_NL), elevation (H_DEM), and land use and land cover (P_LULC). These variables help us to comprehensively capture the various environmental factors affecting atmospheric CO₂ concentrations. In other words, the more frequently a particular attribute is used during the model building, the higher its importance. To verify the accuracy of the OCO-2 remote sensing data, we extracted observational data within a ±5° radius around the Xianghe station of the Total Carbon Column Observation Network (TCCON) for comparative analysis and analyzed the contribution rates of auxiliary variables in the model. In this study, a simple linear regression analysis method was used to analyze the change in correlation relationship significance (p) and slope (b) to analyze the trend of atmospheric x_CO2 concentration from April to August 2023. In addition, a feature importance analysis was conducted to evaluate the significance of each independent variable in the model. The results demonstrated a significant correlation between the OCO-2 satellite remote sensing data and ground-based observational network data, with a coefficient of determination (R²) of 0.902 and a mean squared error (σ_MSE) of 1.45×10⁻⁶. Analysis of feature importance revealed associations between x_CO2 concentration and factors related to the natural environment, meteorological conditions, and anthropogenic activities. Among these factors, the effects of H_DEM and t on the variation in x_CO2 concentration were the most pronounced. This suggests that the distribution of atmospheric CO₂ is significantly regulated by altitude under different topographic conditions, which may be closely related to variations in atmospheric pressure, surface fluxes, and the structure of the atmospheric boundary layer. Changes in temperature directly influence the dynamic balance of carbon sources and sinks at the surface, thereby affecting the spatiotemporal characteristics of x_CO2. In contrast, the contributions of P_LULC and P_t to the variation in x_CO2 were relatively minor. Additionally, while precipitation plays an important role in regulating vegetation growth and soil carbon release, its direct driving effect on x_CO2 concentration was not prominent in this study. Further model analysis showed an R² exceeding 0.82 and an σ_MSE below 0.48×10⁻⁶ between OCO-2 data and RF model predictions, with a mean bias error (E) less than 0.02×10⁻⁶, indicating high model fitting accuracy and strong predictive capability. The spatiotemporal distribution characteristics of atmospheric x_CO2 in China were also analyzed. Temporally, x_CO2 peaked in April and reached its lowest level in August. The concentration trend followed the order: Spring>Winter>Autumn>Summer, exhibiting distinct seasonality. Overall, as temperatures rise in spring, plants enter their growth season, enhancing photosynthesis and fixing significant amounts of atmospheric CO₂. However, this period also sees the melting of winter snow and thawing of soil, which releases carbon from the soil and causes x_CO2 concentrations to rise further. In summer, plants grow vigorously, with intense photosynthesis that fixes large amounts of CO₂. Due to plant transpiration, the air becomes relatively humid, leading to noticeable changes in local x_CO2 concentrations. In autumn, as temperatures drop and daylight hours decrease, plant growth slows, leaves fall, and photosynthesis decreases. The decomposition of fallen leaves and dead branches releases some of the fixed carbon, causing atmospheric x_CO2 concentrations to rise. In winter, low temperatures and shorter days, along with increased heating demand and fossil fuel consumption, lead to a continuous increase in the x_CO2 concentrations. Spatially, atmospheric x_CO2 concentrations generally follow a pattern of being lower in the west and higher in the east, and lower in the north, and higher in the south. The higher the latitude, the greater the seasonal variation, and different temperature zones exhibit variations in the x_CO2 distribution. Based on the analysis of atmospheric x_CO2 concentration spatial distribution changes, it was found that the North China Plain has higher atmospheric x_CO2 concentrations, with the highest levels in Beijing, Tianjin, and Anhui. In Northeast China, atmospheric x_CO2 concentrations exhibit pronounced seasonal changes. From April to August, the significance analysis revealed that, apart from no significant changes in atmospheric x_CO2 concentrations over the Qinghai-Tibet Plateau and the Yunnan-Guizhou Plateau, other regions showed significant variations. Notably, the most significant changes were observed in Heilongjiang, followed by eastern Xizang, Jilin, the Bohai Sea area, and the northern slopes of the Tianshan Mountains in the Xinjiang region. In high-cold ecosystems, even under the backdrop of climate warming, the amount of fixed carbon exceeds the amount released. These ecosystems continue to play a crucial role as carbon sinks, owing to enhanced plant adaptation. This study provides a crucial basis for accurately estimating regional atmospheric CO₂ concentrations in China and offers new perspectives for understanding the carbon cycle processes in terrestrial ecosystems. The research results can help establish a more accurate model, predict the future trend of carbon dioxide change, provide a reference for the refined monitoring of urban carbon emissions, and provide strong geospatial information support for the implementation of the “carbon peak and carbon neutral” strategy at the regional level.

Key words: x_CO2, random forests, spatial and temporal variations, OCO-2, satellite remote sensing

摘要：

为准确把握协同减污降碳和区域碳达峰的战略布局，基于随机森林（RF）模型和轨道碳观测卫星-2（OCO-2）遥感数据，构建2023年中国陆地月尺度时空连续性大气二氧化碳柱平均摩尔分数（x_CO2），空间分辨率为0.1°×0.1°的数据集。使用香河站点的x_CO2数据验证OCO-2观测数据，结果表明两者相关性高，决定系数（R²）为0.902，均方误差（σ_MSE）为1.45×10⁻⁶。选取以自然环境、人为活动、气象条件等影响因素为辅助变量，结合遥感数据训练模型，真实值与预测值之间的R²超过0.82，σ_MSE小于0.48×10⁻⁶，绝对误差（E）小于0.02×10⁻⁶，结果表明模型预测的数据具有极高的可信度。还分析了中国大气CO₂浓度的时空变化分布特征，在时间上，大气CO₂浓度4月达到峰值，8月则降至最低，呈现出明显的季节性变化；在空间上，x_CO2总体呈现“西低东高，北低南高”的空间分布格局，纬度越高，季节性变化越大，不同温度带也表现出x_CO2分布的差异性。该研究对准确估算中国区域大气的x_CO2，以及理解陆地生态系统碳循环的过程至关重要，为城市碳排放工作的精细化监测提供参考，还为区域层面推进“碳达峰、碳中和”战略的实施提供了有力的地理空间信息支撑。

关键词: x_CO2, 随机森林, 时空变化, OCO-2, 卫星遥感