XCO₂ Super-Resolution Reconstruction Based on Spatial Extreme Random Trees

Abstract

Carbon dioxide (CO₂) is currently the most harmful greenhouse gas in the atmosphere. Obtaining long-term, high-resolution atmospheric column CO₂ concentration (XCO₂) datasets is of great practical significance for mitigating the greenhouse effect, identifying and controlling carbon emission sources, and achieving carbon cycle management. However, mainstream satellite observations provide XCO₂ datasets with coarse spatial resolution, which is insufficient to support the needs of higher-precision research. To address this gap, in this study, we integrate spatial information with the extreme random trees model and develop a new machine learning model called spatial extreme random trees (SExtraTrees) to reconstruct a 1 km spatial resolution XCO₂ dataset for China from 2016 to 2020. The results indicate that the predictive ability of spatial extreme random trees is more stable and has higher fitting accuracy compared to other methods. Overall, XCO₂ in China shows an increasing trend year by year, with the spatial distribution revealing significantly higher XCO₂ levels in eastern coastal regions compared to western inland areas. The contributions of this study are primarily in the following areas: (1) Considering the spatial heterogeneity of XCO₂ and combining spatial features with the advantages of machine learning, we construct the spatial extreme random trees model, which is verified to have high predictive accuracy. (2) Using the spatial extreme random trees model, we reconstruct high-resolution XCO₂ datasets for China from 2016 to 2020, providing data support for carbon emission reduction and related decision making. (3) Based on the generated dataset, we analyze the spatiotemporal distribution patterns of XCO₂ in China, thereby improving emission reduction policies and sustainable development measures.

1. Introduction

In recent years, global climate change has become increasingly significant. As carbon dioxide (CO₂) is one of the most harmful greenhouse gases [1,2], CO₂ emissions have become a major factor driving changes in the Earth’s climate system, profoundly impacting its climate. Since the Industrial Revolution, significant amounts of CO₂ have been emitted due to the development of heavy industry and the combustion of large quantities of fossil fuels [3,4], which has caused serious harm to the Earth’s environment and human society [5] and resulted in global warming and an increase in extreme weather events [6,7,8]. Therefore, scholars have extensively explored methods to reduce CO₂ emissions [9,10,11,12], including the development of renewable energy technologies, improving energy efficiency, researching carbon capture and storage methods, and promoting ecosystem restoration. Additionally, recent studies on carbon dioxide (CO₂) emissions assessment have garnered attention [13,14,15]. These studies evaluate the effectiveness of emission reduction policies by examining CO₂ emissions, focusing not only on the technical aspects but also revealing their role in policy implementation and their association with sustainable chemistry. Despite these efforts, global average CO₂ concentrations have continued to increase in recent years. In response to this situation, many governments have enacted greenhouse gas reduction policies and targets [16,17], striving to limit CO₂ concentrations to certain levels. China has also made significant strategic decisions to achieve carbon peaking by 2030 and carbon neutrality by 2060. Therefore, to ensure the realization of these “dual carbon goals” and effectively reduce CO₂ emissions, it is necessary and urgent to accurately understand the spatiotemporal distribution characteristics and trends of CO₂ concentrations.

A surface carbon dioxide monitoring network is a system composed of various ground stations, which is used to monitor the concentration of carbon dioxide (CO₂) in the atmosphere [18]. Examples of such networks include the Global Atmosphere Watch (GAW) led by the World Meteorological Organization (WMO), the Total Carbon Column Observing Network (TCCON) consisting of multiple ground stations worldwide [19,20], the Integrated Carbon Observation System (ICOS) established through collaboration among multiple European countries [21], and the atmospheric monitoring networks under the National Oceanic and Atmospheric Administration (NOAA)’s Earth System Research Laboratory (ESRL), among others. These networks, distributed across multiple ground stations worldwide, monitor the atmospheric CO₂ concentration, providing valuable data on the spatiotemporal distribution of CO₂, which helps improve our understanding of the climate system and carbon cycle. Despite providing important observational data for a period of time, surface CO₂ monitoring networks still have some limitations [22]. For example, the spatial distribution of surface monitoring stations is often sparse and uneven, and the temporal and spatial resolution of the products is low, meaning they are unable to provide high-precision data for the continuous monitoring of atmospheric CO₂ concentrations. This limitation restricts the usability of the data for widespread scientific research.

Satellite remote sensing offers advantages such as wide coverage and long time series, which address the shortcomings of ground-based carbon dioxide monitoring. The development of this technology has provided an important method for accurately monitoring columnar carbon dioxide concentration (XCO₂) and studying the carbon cycle, thereby offering richer data support for global carbon dioxide research [23,24,25,26]. OCO-2 (Orbiting Carbon Observatory-2) and GOSAT (Greenhouse Gases Observing Satellite), launched by NASA and JAXA, respectively, are two of the main satellites providing remote sensing data [27,28,29]. They carry high-resolution infrared absorption spectrometers to measure the absorption spectra of CO₂ in the atmosphere, providing high spatiotemporal resolution data on global CO₂ column concentrations [30]. In addition, TanSat is a CO₂ monitoring satellite launched by China, which utilizes an infrared absorption spectrometer for global CO₂ monitoring [31]. The CO₂ Monitoring Mission planned by the European Space Agency also aims to provide high-quality CO₂ observation data [32]. These satellites, using techniques such as infrared absorption spectroscopy and microwave radiation, have achieved global coverage and relatively high spatiotemporal resolution monitoring, providing a deeper understanding of the spatiotemporal distribution of CO₂ in the atmosphere and its impact on climate change. However, due to limitations in satellite operational cycles, they cannot provide real-time and continuous observational data. Furthermore, although the spatiotemporal resolution of the products has improved compared to previous versions, it still cannot meet the demands of small- to medium-scale carbon environmental monitoring.

Obtaining long-term and high-resolution data on columnar carbon dioxide (CO₂) concentration is crucial for carbon emission monitoring and related research at small to medium scales [33]. However, currently, remote sensing monitoring data with high-precision spatiotemporal resolution are not available. In order to obtain high-precision, high-coverage, and high spatiotemporal resolution CO₂ data products, numerous scholars have conducted research on the high-resolution reconstruction of CO₂ column concentration (XCO₂) using methods such as physical data fusion [34], interpolation, and geostatistics [35,36,37,38]. These interpolation and geostatistical methods model the correlation between known CO₂ points to estimate the concentration at unknown points, providing powerful tools for understanding the spatiotemporal distribution of CO₂. Among them, Kriging interpolation, as a classic geostatistical method, is widely used for estimating spatially continuous surfaces [39,40,41]. Spatial regression analysis combines statistical regression methods with geographic information, thus effectively considering spatial correlations, and provides a more accurate model for interpolation. Additionally, geostatistical cluster analysis divides the study area into spatial units with similar CO₂ concentration characteristics, revealing regional differences in the concentration distribution and providing a reference for regional CO₂ management. Despite the significant progress made as a result of these methods in CO₂ concentration research, they still present some challenges [42]. Geostatistical methods, represented by Kriging interpolation, are sensitive to small-scale spatial changes and cannot effectively handle unstable datasets. They have poor capability in handling the spatial variability of data and can only model and interpolate a single variable, which may ignore significant spatiotemporal correlations between XCO₂ and its influencing factors. Existing statistical models cannot effectively predict nonlinear relationships and are inefficient and less accurate in computing large amounts of data.

Machine learning methods excel at capturing the intricate nonlinear relationships among multiple data sources and efficiently handling large-scale datasets. Therefore, researchers utilize machine learning methods such as artificial neural networks and support vector machines to detect atmospheric pollutants like CO₂ and PM_2.5 [43,44,45]. However, these methods also have their limitations. For instance, parameter tuning in support vector machines significantly affects model performance and requires careful optimization in experiments, while artificial neural network models typically require a large amount of sample data for training. In the downscaling of carbon dioxide concentration, machine learning methods have also been widely applied, and scholars have conducted extensive research to enhance predictive performance [46,47,48]. Methods such as random forest, XGBoost, and gradient-boosting decision trees (GBDTs) have shown excellent applications in the high-resolution reconstruction of carbon dioxide. Random forest is an ensemble learning model that improves model robustness and accuracy by constructing multiple decision trees and combining their results [49]. XGBoost is a gradient-boosting algorithm that enhances the modeling capability of complex spatial relationships and optimizes downscaling effects by handling high-dimensional data, conducting feature selection, and through model optimization [50]. Gradient-boosting decision trees (GBDTs) gradually improve model predictive performance through the iterative training of decision trees [51]. Additionally, extreme random trees (ExtraTrees) have powerful feature dimension reduction capabilities. They introduce randomness in the feature selection and threshold determination process to increase model randomness and diversity, which help to reduce variance and the potential risks of overfitting [52]. By integrating multi-source data, these models handle high-dimensional information, optimize feature selection, enhance the modeling capability of complex spatial relationships, and provide strong support for the high-resolution monitoring of carbon dioxide concentrations. However, current popular machine learning methods rarely consider spatial information between carbon dioxide sources. Solely relying on changes in latitude and longitude values cannot adequately characterize and quantify the spatial position or relationships of geographic features, thus resulting in poor generalization performance of the established XCO₂ prediction models [53].

In summary, the aim of this study was to develop a machine learning model that takes into account spatial characteristics to elucidate the relationship between XCO₂ and its influencing factors. We considered the geographical and temporal heterogeneity of carbon dioxide. During model construction, a geographical spatial encoding strategy was employed to transform the latitude and longitude coordinates of each numerical point into Cartesian coordinates in a three-dimensional Cartesian coordinate system. This strategy ensures the smooth and uniform variation of spatial features to represent the positional relationships of geographic elements, i.e., spatial information. During model construction, spatial information was integrated with extreme random trees, resulting in a machine learning model called spatial extreme random trees (SExtraTrees). This model effectively addresses issues in respect of uneven changes in geographical and temporal location features and nonlinear relationship modeling. The contributions of this study are summarized as follows:

(1)

Incorporating spatial information into the extreme random trees model, which integrates the advantages of machine learning methods and spatial features, enhancing the predictive performance of the model. This model can offer novel approaches for the high-resolution reconstruction of diverse geographical data.

(2)

Utilizing the constructed model to reconstruct a 1 km high-resolution XCO₂ dataset for China from 2016 to 2020.

(3)

Based on the predicted results, identifying and analyzing the spatiotemporal patterns of XCO₂ within the Chinese region. These long-term, high-resolution XCO₂ data contribute to our understanding of the spatiotemporal distribution characteristics of and variations in XCO₂, thus improving the formulation of policies and strategies for carbon emission reduction.

2. Materials and Methods

2.1. Study Area

We aimed to reconstruct the XCO₂ data of the Chinese region (as shown in Figure 1) at high resolution. With a vast land area and diverse geographical environments, China, as the world’s second-largest economy, significantly influences global greenhouse gas emissions. China exhibits diverse terrain, including mountains, plateaus, plains, basins, and hills. Additionally, China’s climate types encompass tropical, subtropical, temperate, and cold temperate regions, showing distinct regional differences. These topographical and climatic features have significant impacts on the generation, transport, and distribution of carbon dioxide. For instance, mountainous regions may have lower carbon dioxide concentrations, while plain areas, due to dense populations and frequent industrial activities, may exhibit higher carbon dioxide emissions. Therefore, by conducting a more in-depth study to understand the spatial distribution and temporal trends of carbon dioxide concentrations in China, we can gain better insights into carbon cycle processes.

2.2. XCO₂ Data

The XCO₂ data used in this study were derived from a 2016–2020 global coverage XCO₂ dataset generated by integrating multiple sources, including satellite data and reanalysis data [54]. Specifically, the dataset was generated by establishing the relationship between OCO-2 satellite XCO₂ retrieval data and Copernicus Atmosphere Monitoring Service (CAMS) XCO₂ reanalysis data with influencing factors such as NPP and RH using deep neural networks (DNNs). Subsequently, comprehensive XCO₂ data were generated using CAMS XCO₂ data and associated influencing factors. In this study, XCO₂ data were modeled at a spatial resolution of 10 km. Therefore, following quality control, any null values and outliers in the dataset were removed to form the final XCO₂ dataset at a resolution of 10 km.

2.3. Influencing Factors

Studies have shown that there are many factors that influence carbon dioxide concentration, among which population density (POP), temperature (TEM), normalized difference vegetation index (NDVI), precipitation (PRE), and gross primary productivity (GPP) are important factors affecting columnar carbon dioxide concentration (XCO₂) [55,56]. In this study, we classified influencing factors into three categories: human activities, vegetation, and meteorological factors [57]. Areas with higher population density exhibit significantly higher carbon dioxide concentrations, as the distribution of emission sources such as industry and transportation directly affects carbon dioxide concentrations [58,59]. Vegetation conditions directly affect the absorption of carbon dioxide by vegetation and also influence the exchange processes between vegetation and the atmosphere. Meteorological factors such as temperature and precipitation also have significant effects on carbon dioxide concentrations. Therefore, in this study, factors related to human activities included population density (POP) and nighttime light index (NTL), vegetation conditions include gross primary productivity (GPP) and normalized difference vegetation index (NDVI), and meteorological conditions included temperature (TEM), precipitation (PRE), humidity (RH), and wind speed (WND). Additionally, we considered the influence of topography on carbon dioxide column concentration and incorporated digital elevation models (DEMs) for modeling purposes.

Population density has a significant impact on carbon dioxide (CO₂) concentration. Regions with high population density are often associated with increased urbanization, industrialization, and commercial activities, which may lead to higher energy consumption and industrial emissions, thus increasing CO₂ emissions [60,61]. The LandScan dataset used in this study is developed and maintained by the Oak Ridge National Laboratory (ORNL) of the Department of Energy (DOE) in the United States. It combines high-resolution satellite imagery, census data, geographic information system (GIS) data, and other geographical, topographical, and meteorological data, generated through complex models and algorithms, with a spatial resolution of 1 km [62]. The nighttime light (NTL) index is typically calculated using nighttime light images obtained from remote sensing satellites [63]. Its primary applications include urban planning, environmental monitoring, and economic development research. By measuring and recording the brightness of the Earth’s surface, nighttime light images can reveal the population density, level of economic activity, and urban development of an area. A high NTL index may indicate a higher level of economic development in the area, accompanied by increased energy use and CO₂ emissions [64]. The NTL data used in this study were obtained from the VIIRS sensor on the Suomi NPP satellite and processed to eliminate noise from different dimensions to obtain the final dataset.

NDVI can accurately reflect the vegetation cover status of the Earth’s surface, and previous studies have shown a significant negative correlation between NDVI and carbon dioxide concentration [65]. We utilized an annual vegetation index dataset based on uninterrupted time series derived from SPOT/VEGETATION NDVI satellite remote sensing, with a spatial resolution of 1 km [66]. The NDVI data were sourced from the Resource and Environmental Science Data Center [67]. Gross primary productivity (GPP) is commonly used to describe the total amount of photosynthesis performed by plants in terrestrial ecosystems. Photosynthesis is the process by which plants utilize sunlight to convert carbon dioxide and water into organic substances. This process enables plants to capture and store carbon dioxide from the atmosphere, making it crucial for the carbon cycle in the atmosphere. Therefore, monitoring and studying GPP can provide a better understanding of the absorption, fixation, and release of carbon by plant ecosystems, which is important for understanding carbon dioxide concentration. The GPP data used in this study are sourced from the Global Ecology Group website and have a spatial resolution of 5 km [68].

Meteorological factors, such as temperature and precipitation, may impact the carbon cycle of terrestrial ecosystems and soil respiration [69]. Humidity impacts plant transpiration, affecting the rate at which plants absorb carbon dioxide [70]. Wind can disperse local gases, resulting in relatively lower carbon dioxide concentrations in localized areas [71]. The precipitation and temperature data used in this study are sourced from publicly available datasets at the National Earth System Science Data Center [72], while the humidity and wind speed datasets are from the Resource and Environmental Science Data Center [73]. The detailed information and sources of the selected independent variables are shown in

Table 1. Independent variables and corresponding sources.

Variable	Spatial Resolution	Source	Website
DEM	1 km × 1 km	Resource and Environmental Science Data Center	https://www.resdc.cn (accessed on 10 February 2024)
GPP	5 km × 5 km	Global Ecology Group	https://globalecology.unh.edu/ (accessed on 10 February 2024)
NDVI	1 km × 1 km	Resource and Environmental Science Data Center	https://www.resdc.cn (accessed on 10 February 2024)
NTL	1.5 km × 1.5 km	Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences	https://www.zybuluo.com/novachen/note/1162587 (accessed on 10 February 2024)
POP	1 km × 1 km	Oak Ridge National Laboratory	https://landscan.ornl.gov/ (accessed on 10 February 2024)
WND	1 km × 1 km	Resource and Environmental Science Data Center	https://www.resdc.cn (accessed on 10 February 2024)
PRE	1 km × 1 km	National Earth System Science Data Center	http://loess.geodata.cn/ (accessed on 10 February 2024)
TEM	1 km × 1 km	National Earth System Science Data Center	http://loess.geodata.cn/ (accessed on 10 February 2024)
RH	1 km × 1 km	Resource and Environmental Science Data Center	https://www.resdc.cn (accessed on 10 February 2024)

.

2.4. Data Processing

In this study, our aim was to estimate the XCO₂ in China at a spatial resolution of 1 km annually from 2016 to 2020. Therefore, we employed the nearest-neighbor interpolation method to resample all the aforementioned influencing factor data, ensuring they had a consistent spatial resolution of 1 km. Subsequently, we conducted quality control on the resampled influencing factor dataset, removing outliers and missing values. Then, we aggregated the quality-controlled 1 km influencing factor data by year, resulting in five 1 km input datasets for model prediction from 2016 to 2020. Next, we projected the 1 km influencing factor datasets onto the WGS84 projection coordinate system, resampled the 1 km resolution factors to a 10 km resolution using the nearest-neighbor method, and spatially matched the results with the original resolution of the XCO₂ data (i.e., 10 km × 10 km) to generate training sample datasets. We obtained a total of 469,685 sample data points during the period from 2016 to 2020, with a spatial resolution of 10 km. Subsequently, we conducted modeling using the processed training sample dataset.

2.5. Spatial Extreme Random Trees

China is a vast territory, with significant differences in urbanization levels, industrial conditions, energy structures, and climate conditions among different regions, resulting in typical spatial distribution patterns of carbon dioxide concentration. There is evident spatial heterogeneity between carbon dioxide concentration and influencing factors. Although the extreme random trees method can increase the randomness and diversity of models and reduce the risk of overfitting by randomizing feature selection and threshold selection processes, better capture the nonlinear relationships between data, and further improve the robustness of the model, it cannot effectively address the spatial heterogeneity of XCO₂. Therefore, by considering spatial features and incorporating spatial information into the extreme random trees model, we can inherit the advantages of extreme random trees and effectively address the uneven changes in geographical and temporal location characteristics. In this study, by adopting a geographic spatial encoding strategy (Equation (1)), latitude and longitude polar coordinates were converted into Cartesian coordinates in a three-dimensional Cartesian coordinate system to characterize the positional characteristics of geographical elements (Figure 2), resulting in feature variables that uniformly change within the range [−6371, 6371], i.e., spatial relationships. This spatial information exhibits smooth and uniform changes, effectively addressing the spatial heterogeneity issue of XCO₂.

$$g_i=\left[\begin{array}{c}{g}_{xi}\\ g_{yi}\\ g_{zi}\end{array}\right]=\left[\begin{array}{c}Rsin{\phi }_{i}cos{\theta }_i\\ Rsin{\theta }_i\\ Rcos{\phi }_{i}cos{\theta }_i\end{array}\right]$$

(1)

In the equation, $g_i$ represents the geographic spatial information encoding feature of point i, ${\phi }_i$ represents the longitude of point i,${\theta }_i$ represents the latitude of point i, and R represents the radius of the Earth (approximately 6371 km).

The spatial extreme random trees model was used to establish the relationship between XCO₂ and influencing factors at a resolution of 10 km. The specific principles of the spatial extreme random trees are illustrated in Figure 3. Initially, the pre-processed annual scale environmental factor data set with a spatial resolution of 1 km from 2016 to 2020 was resampled to an annual scale data set with a spatial resolution of 10 km by the nearest neighbor method, and aligned with the original XCO₂ annual scale data set with a spatial resolution of 10 km. The XCO₂ data of each point has corresponding environmental factor data, which establishes the corresponding relationship between the model-independent variables and some dependent variables, and this process is called data fusion. Subsequently, spatial information computed based on latitude and longitude was aggregated into the aligned dataset to form a training dataset. The model computation for the relationship between independent variables and dependent variables based on the spatial extreme random trees is represented by Equation (2). XCO₂ data serve as the dependent variable, while factors such as DEM, GPP, and spatial features are used as independent variables for modeling. Leveraging this established relationship, the dataset of influencing factors at 1 km resolution was inputted into the trained spatial extreme random trees model to reconstruct China’s high-resolution (1 km) XCO₂ dataset.

$${XCO}_2=SExtraTrees\left(DEM,GPP,NDVI,NTL,POP,Wnd,Pre,Tem,Rh,Spatial Features\right)$$

(2)

To validate the performance of the spatial extreme random trees method in reconstructing XCO₂, we selected mainstream downscaling methods such as Kriging interpolation and extreme random trees for comparison and analyzed the results of resolution reconstruction using various methods. We employed the 10-fold cross-validation method to evaluate the performance of the spatial extreme random trees. The original dataset was randomly partitioned into ten equally sized subsets, nine of which were allocated for training the model, while the remaining segment was designated for validating the model’s performance. The best model parameters were determined through ten-fold cross-validation. The performance of the established models was evaluated using three evaluation metrics: mean absolute error (MAE), root-mean-squared error (RMSE), and coefficient of determination (R²). The calculation of these statistical metrics can be achieved as follows:

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}$$

(3)

$$\mathrm{MAE}=\frac{1}{n}{\sum }_{i=1}^n|y_i-\widehat{y_i}|$$

(4)

$${\mathrm{R}}^2=1-\frac{{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{{\sum }_{i=1}^n{(y_i-\overline{y_i})}^2}$$

(5)

In this context, $n$ stands for the sample size; $y_i$ and $\widehat{y_i}$ denote the observed XCO₂ value and the predicted XCO₂ value for sample i, respectively; and $\overline{y_i}$ represents the average of the observed XCO₂ values.

3. Results

3.1. Model Comparison

To assess the effectiveness of the proposed model, we compared the predictive results of the spatial extreme random trees model with those of Kriging interpolation and extreme random trees model predictions. The fitting results of each model are shown in

Table 2. Comparison of predictive performance with different methods.

Model	R2	MAE (ppm)	RMSE (ppm)
Kriging	0.961	0.697	0.699
ExtraTrees	0.939	0.726	0.866
SExtraTrees	0.985	0.407	0.434

. Kriging interpolation, as the most classical interpolation method, infers values at unknown locations by leveraging the spatial correlation between known observation points. Previous studies have demonstrated its effectiveness in CO₂-related research [37]. Since the data initially used for this method are XCO₂ data with a resolution of 10 km, and the known points are evenly distributed, the fitting results are relatively good, with an R² of 0.961, MAE of 0.697 ppm, and RMSE of 0.699 ppm. However, this method can only model and interpolate single variables, which ignores the spatiotemporal correlation between XCO₂ and other influencing factors. It may oversmooth the data, fail to capture local variability, exhibit a poor ability to predict nonlinear relationships, and have low efficiency and accuracy when computing large amounts of data. Although the extreme random trees model considers the complex relationships between XCO₂ and other influencing factors and performs well in capturing nonlinear relationships and improving efficiency when processing large amounts of data, it ignores the spatial heterogeneity of XCO₂. The fitting results for the extreme random trees model are as follows: R² = 0.939, MAE = 0.726 ppm, and RMSE = 0.866 ppm. In contrast, the spatial extreme random trees model demonstrates strong nonlinear modeling capabilities. When spatial features were integrated into the machine learning method, this model’s ability to capture spatial geographical relationships was enhanced, leading to significantly improved fitting accuracy with an R² of 0.985, MAE of 0.407 ppm, and RMSE of 0.434 ppm. Hence, we opted to employ spatial extreme random trees for the high-resolution reconstruction of XCO₂.

Figure 4 illustrates the spatial distribution of the initial XCO₂ data at a 10 km resolution in 2016, along with the reconstructed XCO₂ results at a 1 km resolution using different methods. Because Kriging interpolation infers the values at unknown points based on the numerical values of known points, and XCO₂ generally exhibits a gradual trend, visually, the Kriging interpolation results show a clear smoothing trend. By ignoring spatial heterogeneity, the extreme random trees model theoretically underestimates the complexity of XCO₂ spatial distribution, leading to the phenomenon of underestimation in the predicted XCO₂ values, particularly in eastern coastal areas. The spatial distribution of XCO₂ predicted by the spatial extreme random trees model is most similar to the original data, indicating that our reconstructed XCO₂ data possess higher precision and accuracy.

3.2. Spatial Extreme Random Trees Validation

In this study, based on the spatial extreme random trees model, the 10 km-resolution XCO₂ spatiotemporal dataset for China from 2016 to 2020 was reconstructed at a 1 km resolution. The model fitting results for each year are depicted in Figure 5. The fitting R² values range from 0.949 (2017) to 0.964 (2018), indicating high precision in both RMSE and MAE results. The accuracy of the prediction results demonstrates that the spatial extreme random trees model can reliably predict XCO₂ at relatively fine resolutions.

3.3. Spatiotemporal Variations in XCO₂ in China

The results of China’s 1 km-resolution XCO₂ spatiotemporal dataset reconstructed using spatial extreme random trees from 2016 to 2020 are shown in Figure 6, indicating notable spatial variations in the distribution of XCO₂. Overall, consistent with the existing research results, the carbon dioxide concentration tends to be relatively lower in western China and higher in eastern China, a pattern that aligns with the level of urbanization and economic development in China [74]. The rapid urbanization in China from 2016 to 2020 has led to significant differences in XCO₂ between developed eastern regions and underdeveloped western regions. The eastern regions of China are typically characterized by economic prosperity, dense population, and intensive energy and industrial activities, which may result in higher carbon dioxide concentrations. Cities such as Beijing, Shanghai, and parts of Jiangsu province often exhibit higher population densities, traffic volumes, and industrial activities, leading to higher carbon dioxide concentrations in urban areas. In contrast, the western and northeastern regions of China may be more influenced by natural factors. Regions with high vegetation coverage, such as Tibet, Inner Mongolia, and Heilongjiang, absorb more carbon dioxide through plant photosynthesis, resulting in lower carbon dioxide concentrations. Additionally, the western regions have lower population densities and are gradually transitioning from traditional to new energy sources (such as wind and solar energy), thus reducing carbon emissions. The economy in the southwestern and southern regions is relatively developed, with extensive human activities. However, because of the warm and humid climate prevailing in these regions, vegetation is lush, resulting in the higher efficiency of carbon fixation through photosynthesis. Therefore, the XCO₂ level is moderate [75]. Further, optimizing the energy structure and improving energy efficiency to decrease carbon intensity, as well as increasing vegetation coverage to enhance carbon sequestration, are effective approaches for reducing or mitigating XCO₂. The 1 km-resolution XCO₂ data reconstructed via spatial extreme random trees can achieve comprehensive coverage with high resolution in China, thereby enabling more effective monitoring of carbon sources and sinks.

The spatial distribution of XCO₂ trends in China from 2016 to 2020, as shown in Figure 7, is consistent with the global trend, with an overall increase in XCO₂ levels nationwide each year [76]. Specifically, all regions exhibit significant increments, with an average annual rise of over 1.88 ppm/year over the past five years. The rate of XCO₂ growth varies across regions. In the northwest, northeast, and northern parts of China, the concentration of carbon dioxide in the atmosphere has increased rapidly, with XCO₂ levels higher than the five-year average. This could be attributed to accelerated urbanization in China, intensive human activities, and increased energy consumption that have led to higher carbon emissions. In contrast, although XCO₂ levels in southwest, central, eastern, and southern China have increased annually, the growth rates have been relatively slow. For example, XCO₂ levels increased more slowly near cities such as Chengdu, Chongqing, Changsha, and Wuhan. This may be due to the rapid economic development, high population density, and high level of urbanization in these regions, which emit more carbon dioxide, resulting in a large carbon dioxide concentration base and in a lower growth rate in these regions. Additionally, the warm and humid climate in the southern regions supports lush vegetation, which efficiently absorbs carbon dioxide through photosynthesis, mitigating the increase in atmospheric carbon dioxide. Moreover, eastern and southern regions of China have gradually transitioned from energy-intensive economic growth to green and sustainable development, which may have contributed to the slowdown in XCO₂ growth rates. When studying the spatiotemporal distribution of XCO₂, it is crucial to consider the non-uniformity of carbon dioxide emissions. Therefore, obtaining long-term, high-resolution XCO₂ data can provide a scientific basis for China to formulate reasonable carbon emission reduction policies. To better assess carbon emissions within China, reasonable CO₂ emission analyses are needed across different regions.

4. Conclusions

As CO₂ is a major greenhouse gas, excessive CO₂ emissions are a significant contributor to global warming. In this study, we introduced a novel spatial downscaling model named SExtraTrees, aimed at addressing the research gap in reconstructing high-resolution XCO₂ from the perspective of spatial heterogeneity. By considering nonlinear relationships and spatial characteristics, SExtraTrees further enhances the accuracy and robustness of high-resolution XCO₂ reconstruction, thereby improving the model’s generalizability. Based on XCO₂ data with a spatial resolution of 10 km from 2016 to 2020, we successfully reconstructed XCO₂ at a spatial resolution of 1 km. To meet the need for carbon emission monitoring, the spatiotemporal characteristics of XCO₂ were analyzed within the region of China. Based on the results of this study, the following conclusions were drawn:

(1)

The SExtraTrees model proposed in this study is an effective technique for predicting national long-term time series XCO₂ data. The SExtraTrees model outperforms Kriging interpolation and extreme random trees models in terms of model fitting and prediction while maintaining spatial consistency with the original XCO₂ data and ensuring high predictive accuracy.

(2)

The 1 km-resolution XCO₂ products reconstructed using the spatial extreme random trees model provide fundamental data and technical support for regional XCO₂ monitoring. These predicted XCO₂ datasets can also serve as a reference for calibrating other low-resolution XCO₂ datasets. The obtained results highlight the importance of combining geographic spatial correlations with machine learning methods to achieve high precision and robustness. The spatial extreme random trees model can also be applied to the resolution reconstruction of other environmental factors.

(3)

Based on the resolution reconstruction results of the spatial extreme random trees model, XCO₂ exhibits significant spatial and temporal heterogeneity. From 2016 to 2020, XCO₂ in China shows an increasing trend each year. Nationally, the spatial distribution of XCO₂ aligns with China’s economic development and urbanization level, with higher XCO₂ concentrations in the eastern regions and lower concentrations in the western regions of China. The distribution of carbon concentrations varies among different regions, mainly due to differences in the geographical environment, economic development level, and industrial structure. Higher vegetation coverage can enhance the carbon sequestration capacity. Given the spatiotemporal changes and differences in XCO₂, this study provides scientific references for reducing carbon emissions, which is essential for crafting efficient policies aimed at reducing carbon emissions and addressing climate change.

In summary, the spatial extreme random trees model enhances the fitting capability of XCO₂ with complex spatial features and exhibits good generalization performance. SExtraTrees may also be applicable to other studies related to high-resolution reconstruction of air pollutants, serving as a methodological reference for reconstructing other factors at high resolutions. It is worth noting that in other relevant studies, when establishing models, it is essential to select influential factors that are strongly correlated with the dependent variable. Through the application of the spatial extreme random trees model, we have gained deeper insights into the spatiotemporal distribution characteristics of XCO₂ in China, providing robust data support for future carbon emission control and climate change research. In future studies, it will be necessary to further consider the unique temporal characteristics of XCO₂ data. Additionally, given the diverse and complex environmental factors influencing XCO₂, there is a necessity for more in-depth research into the intricate relationships between influencing factors and XCO₂. By employing spatial statistical analysis and time series analysis, we aim to thoroughly explore the spatiotemporal information embedded in long time series CO₂ satellite observations, meteorological station data, and other relevant datasets. By integrating data from various time points and spatial locations, we seek to acquire a more comprehensive and holistic understanding of spatiotemporal dynamics, thereby continuously enhancing the resolution reconstruction models for XCO₂ data.