Mapping High-Resolution Carbon Emission Spatial Distribution Combined with Carbon Satellite and Muti-Source Data

Abstract

Carbon satellites, as the most direct means of observing carbon dioxide globally, offer credible and scientifically robust methods for estimating carbon emissions. To enhance the accuracy and timeliness of urban-scale carbon emission estimates, this study proposes an innovative model that integrates top-down carbon satellite data with high-resolution spatial proxies, including points of interest, road networks, and population distribution. The K-means clustering method was employed to study the relationship between carbon emissions and XCO₂ anomalies. Based on this, the local adaptive carbon emission estimation model was constructed. Further, by integrating the spatial distribution and weights of proxy data, carbon emissions were reallocated to generate a high-resolution urban carbon emission map at a 1 km × 1 km resolution. Taking Urumqi, the XCO₂ background concentration ranged from approximately 408 ppm to 415 ppm in 2020, and the corresponding ${∆\mathrm{XCO}}_2$ ranged from −1.58 ppm to 1.13 ppm. The total carbon emission estimated by the local adaptive model amounted to approximately 58.26718 million tons in 2020, close to the EDGAR dataset, with most monthly relative error within ±10%. The Pearson correlation coefficient between the ODIAC dataset and spatially redistributed carbon emission was 0.192, and their comparison showed that high carbon emission areas in the spatially redistributed carbon emission aligned closely with urban industrial parks and commercial centers, offering a more detailed representation of urban carbon emission spatial characteristics. This method contributed to exploring the potential of carbon satellites for quantitatively measuring anthropogenic emissions and offers improved insights into monitoring urban-scale carbon dioxide emissions.

1. Introduction

Reducing carbon emissions to mitigate global warming remains one of the world’s most pressing tasks [1]. Research on carbon emission estimation has increasingly gained attention. Traditionally, most carbon emission estimates have used administrative unit-level production statistics, such as China’s provincial carbon emission inventories and Emissions Database for Global Atmospheric Research (EDGAR), following methodologies developed by the Intergovernmental Panel on Climate Change (IPCC) or provincial guidelines [2,3]. While these approaches are highly credible, they fail to capture spatial variations and energy consumption patterns within administrative units, limiting their applicability for interdisciplinary actions aimed at addressing climate change. Spatial modeling of carbon emissions offers a solution by enabling the development of spatially distributed emission inventories and uncovering spatial patterns within administrative boundaries [4,5,6].

Currently, the main approaches for estimating the spatial distribution of carbon emissions include top-down, bottom-up, and hybrid methods combining both [7]. The top-down approach mainly uses nighttime light data and population data to estimate carbon emissions. While this method provides medium accuracy (70–90%) and typically offers monthly or annual temporal resolution, its spatial resolution can reach up to 1 km. However, due to saturation effects in nighttime light data, the results may fail to reflect the intra-urban emission pattern accurately. Therefore, this method is widely applied at regional and national scales, with limited application in city-level studies [8,9]. The bottom-up approach usually uses carbon emission inventories for statistical analysis at the city, regional, or community scale. This method has a high spatial resolution and the most accurate estimates [10,11,12]. However, it has certain limitations. On the one hand, due to the complex calculation process of this method and the involvement of a lot of non-public data, it has a time lag and is usually an annual statistic. On the other hand, it is difficult to obtain carbon emission statistics in underdeveloped regions, so it is also a problem to accurately estimate carbon emissions. The combined top-down and bottom-up approach has the strengths of both methods, providing improved accuracy and spatial detail. Consequently, it has gained increasing adoption in recent studies [13,14].

In recent years, with the increasing availability of carbon satellite observation data, researchers have gradually begun using this data to estimate anthropogenic carbon emissions [15,16]. Compared to the bottom-up approach, this method offers higher temporal resolution and greater timeliness, enabling estimates on weekly, monthly, or annual scales. Additionally, the data is publicly accessible. However, the primary challenge lies in the missing accuracy of the estimated results, and only a few studies have been conducted on carbon emissions estimation in large power plants and large industrial areas [17,18]. The OCO-2 satellite offers the finest spatial resolution with a 1–2 times higher order of magnitude (1.25 × 2 km) and a promising measurement accuracy (0.5–1 ppm) for city-scale research [19]. As a result, studies using OCO-2 observations to attribute CO₂ signals to urban emissions have shown promising progress [20,21]. This advancement provides the possibility of further using carbon satellite observations to refine the spatial distribution of carbon emission estimates.

In order to map the spatial distribution of carbon emissions at a more refined local scale, many studies have used high-resolution spatial proxy data. For example, ref. [22] combined the geographic information of points of interest (POI) and night lights to estimate carbon emissions and mapped the distribution of urban carbon emissions at a resolution of 130 m. Wu, et al. [23] evaluated the spatiotemporal patterns of transportation carbon emissions using taxi GPS trajectory data and points of interest and road network data. Liu, et al. [24] used points of interest, road networks, and land use to determine the location of emission sources and allocate emissions to map high-resolution carbon emission distribution maps. Spatial proxy data plays a role in approximating the spatial allocation of entity-level in the estimation of carbon emission distribution, which can improve the spatial resolution of carbon emission distribution maps.

The estimation of carbon emissions, conducted by directly observing carbon dioxide concentration data in the Earth’s atmosphere using carbon satellites, and further utilizing proxy data to allocate carbon emissions, provides a relatively comprehensive, accurate, and timely inversion of the high-resolution spatial distribution of carbon emissions. In this study, we focus on Urumqi, China, using XCO₂ observations from the OCO-2 satellite measurement data and spatial proxy data such as POI, population density, and road distribution. After preprocessing the dataset, the local adaptive model was constructed to estimate the carbon emission map at 10 km spatial resolution. Subsequently, a spatial redistribution model was applied to further downscale the results to a spatial resolution of 1 km and a monthly temporal resolution.

2. Study Area and Data

2.1. Study Area

Urumqi is located in the Urumqi Valley at the northern foot of the Tianshan Mountains and the southern edge of the Junggar Basin. The city is surrounded by steep mountains and hills on nearly three sides. The average altitude of the urban area is 800 m, see Figure 1. Urumqi is characterized by a temperate continental arid climate with distinct seasons. Spring and autumn are relatively short, occurring from late March to early June and from late August to early November, respectively. In contrast, summer and winter are prolonged, lasting from early June to late August and from early November to late March of the following year. The average annual precipitation is approximately 294 mm. The warmest months are July and August, with an average temperature of 25.7 °C, while the coldest month is January, with an average temperature of −15.2 °C.

With the rapid economic growth in the past two decades, the number of industrial plants such as power plants, petrochemical plants, and cement plants has increased, and the number of motor vehicles has increased rapidly, so the city’s carbon emissions have increased sharply. The carbon emissions of different types in Urumqi vary greatly, among which industrial carbon emissions account for a large proportion. It is necessary to draw a carbon emission spatial distribution map of 1 km grid according to different types of carbon emission sources such as commercial, industrial, transportation, and residential.

2.2. OCO-2 XCO₂ Data

The XCO₂ data used in this study were obtained from the Orbiting Carbon Observatory-2 (OCO-2), a satellite launched by the National Aeronautics and Space Administration (NASA) specifically for global carbon dioxide monitoring. OCO-2 measures column-averaged CO₂ dry air mole fractions (XCO₂) at a native spatial resolution of 2.25 km × 1.29 km, with a revisit cycle of 16 days.

The data used in this study is the OCO2_GEOS_L3CO2_MONTH version, with a spatial resolution of 0.5° × 0.625° and a monthly temporal resolution. This Level 3 product is generated by assimilating Level 2 XCO₂ retrievals into the GEOS data assimilation system and includes comprehensive atmospheric corrections such as for aerosols, clouds, and surface pressure.

This study downloaded the monthly data for 2020 and used the bilinear method to interpolate the data into a data set with a spatial resolution of 10 km. The data can be downloaded from https://disc.gsfc.nasa.gov/ (accessed on 20 October 2024).

2.3. Spatial Proxy Data

Spatial proxy data refers to the indirect acquisition of spatial data that is difficult to measure directly or cannot be obtained through alternative data. The spatial proxy data used in this study include POI data, road network data, and population density data, which can reflect the size of carbon emissions to a certain extent and are used to estimate carbon emissions in maps with smaller scales and higher spatial resolutions.

POI data refers to data used to mark a specific place or location, usually including a description, category, and coordinates of the place. This study obtained the POIs of Urumqi in 2020 through the Amap open interface (https://lbs.amap.com/api/, accessed on 15 June 2024), retaining five industry labels: catering, shopping, hotel accommodation, and life services, of which the “factories” type was used as a separate category. Road network data comes from OpenStreetMap (https://www.openstreetmap.org/export, accessed on 15 June 2024). Motorway, Trunk, Primary, Secondary, Tertiary, and Residential roads were retained during data cleaning, and other road types were ignored. Population data comes from LandScan Global (https://landscan.ornl.gov/, accessed on 11 September 2024), which provides a spatial distribution map of yearly population numbers with a spatial resolution of 30 arc-seconds (about 1 km). A detailed description of POI and Road data is summarized in

Table 1. Spatial proxy data description.

Data Source	Category Label	Count	Spatial Proxy Function
POI	catering, shopping, hotel accommodation, life services	70173	Carbon emission space allocation for services.
POI	factories	281	Carbon emission space allocation for industry.
Road	motorway, trunk, primary, secondary, tertiary, residential roads	-	Carbon emission space allocation for road traffic.
Population	population	-	Carbon emission space allocation for daily consumption.

.

2.4. Global Gridded Emission Datasets

The study used two carbon emission datasets, Emissions Database for Global Atmospheric Research (EDGAR) and Open-source Data Inventory for Anthropogenic CO₂ (ODIAC), for model construction and verification of carbon emission estimates, respectively [25].

EDGAR is a global database of anthropogenic greenhouse gases constructed using a bottom-up carbon emissions inventory approach. Emissions were calculated by using a technology-based emission factor approach consistently applied for all world countries. The spatial allocation of emissions is in grid cells of 0.1 degrees by 0.1 degrees. A geographical database was built using spatial proxy datasets with the location of energy and manufacturing facilities, road networks, shipping routes, human and animal population density and agricultural land use, that vary over time. This study used EDGAR to study the relationship between anomalies of XCO₂ and carbon emissions, and fused EDGAR and observational XCO₂ data. The data download address is: https://edgar.jrc.ec.europa.eu/emissions_data_and_maps (accessed on 20 October 2024).

ODIAC is a global high-resolution fossil fuel CO₂ emission (ffco2) data product. The emission spatial disaggregation was done using multiple spatial proxy data, such as geographical location of point sources, satellite observations of nightlights, and aircraft and ship fleet tracks. This study uses the ODIAC2022 version 1 km by 1 km data to verify results. The data download address is: https://db.cger.nies.go.jp/dataset/ODIAC/ (accessed on 20 October 2024). What needs attention is the difference between the carbon emission estimation methods of ODIAC and EDGAR. ODIAC adopts a top-down approach, while EDGAR adopts a bottom-up approach.

3. Methodology

This study integrated top-down OCO-2 XCO₂ data with bottom-up spatial proxies to estimate high-resolution urban carbon emissions. The methodological framework consists of four main components: (1) calculation of XCO₂ background concentration and XCO₂ anomalies, (2) spatiotemporal clustering and local adaptive modeling, (3) spatial redistribution of carbon emission estimation, and (4) accuracy assessment and validation. The overall road map is illustrated in Figure 2.

3.1. Calculation of XCO₂ Background Concentration and XCO₂ Anomalies

XCO₂ background concentration refers to the XCO₂ concentration value in an area that is almost unaffected by local human activities. XCO₂ anomalies refer to the difference between the XCO₂ value and the XCO₂ background concentration. In the following text, $∆{\mathrm{XCO}}_2$ is used to represent XCO₂ anomalies. Previous studies have shown that $∆{\mathrm{XCO}}_2$ can de-seasonalize variations caused by seasonal changes and reflect point source information of carbon emissions [16]. This study obtained areas that can represent the background concentration of XCO₂ through the overlay analysis of spatial proxy data such as POI, and calculated the mean of these areas as the background concentration of this study, thereby further obtaining $∆{\mathrm{XCO}}_2$. The specific background concentration value determination process is as follows:

(1)

Perform kernel density analysis on POI and Roads, where Road sets different reference weights of 1–5 according to different road levels [26]. Overlay different kernel density analysis results and generate a map, which can reflect the degree of human activity.

(2)

Perform spatial aggregation on population, aggregate 1 km data into a 10 km grid map, and select areas with a population of 0 to generate a map. The population density is closely related to human activities.

(3)

Overlay the above kernel density superposition results with the population map to obtain an area that is almost unaffected by human activities. Generate sampling points within this area and extract the mean XCO₂ concentration as the monthly XCO₂ background concentration value.

The difference between the spatial map of XCO₂ concentration values and the background concentration value for each month is $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$.

3.2. Spatiotemporal Clustering and Local Adaptive Modeling

To establish a reliable relationship between carbon emissions from EDGAR and $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$ from OCO-2, this study adopted spatiotemporal clustering analysis to capture the variation patterns of both datasets across different spatial locations. Based on the identified patterns, the local adaptive model was subsequently developed to enhance the spatial integration of the two datasets.

(1) Spatiotemporal clustering. The $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$ and EDGAR carbon emission were resampled to a spatial resolution of 10 km using bilinear interpolation. For each grid cell, time series of the $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$ and EDGAR carbon emission over 12 months were extracted. Subsequently, the K-means clustering method was applied to identify the spatiotemporal variation patterns.

For a given time series $\mathrm{X}=\{{\mathrm{x}}_1,{\mathrm{x}}_2,{\mathrm{x}}_3,\dots ,{\mathrm{x}}_{12}\}$, the algorithm flow of K-mean clustering is as follows:

Step 1. Initialize the centroid: randomly select K data points as the initial centroid;

Step 2. Assign data points: assign each data point to the cluster to which the nearest centroid belongs;

Step 3. Update the centroid: recalculate the centroid of each cluster, and the new centroid is the average of all data points in the cluster.

Repeat steps 2 and 3 until the centroid no longer changes.

Before the algorithm is executed, it is necessary to standardize the data of EDGAR carbon emissions and $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$, and finally, through repeated iterations, the objective function converges to the local minimum.

(2) Local adaptive modeling. Due to the heterogeneity of geographic spatial relationships, it is difficult to construct the relationship between $∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2$ and EDGAR carbon emission through a specific linear function. Therefore, a local adaptive regression model is developed within each grid cell to quantify the relationship between the two, based on the optimal functional form identified for that unit [27]. The model is described as follows:

A set of candidate functional forms—linear, binomial, logarithmic, and exponential—were evaluated, with the mean squared error (MSE) used as the selection criterion [28]. The selected optimal relationship for each unit is expressed as follows:

$$\left\{\begin{array}{c}\mathrm{emission}=\mathrm{a}\times {∆\mathrm{X}\mathrm{C}\mathrm{O}}_2+\mathrm{b}\\ \mathrm{emission}=\mathrm{a}\times {({∆\mathrm{X}\mathrm{C}\mathrm{O}}_2)}^2+\mathrm{b}\times ∆{\mathrm{X}\mathrm{C}\mathrm{O}}_2+\mathrm{c}\\ \mathrm{emission}=\log ({∆\mathrm{X}\mathrm{C}\mathrm{O}}_2+\mathrm{k})\\ \mathrm{emission}=\exp ({∆\mathrm{X}\mathrm{C}\mathrm{O}}_2)\end{array}\right.$$

where emission refers to the estimated value of carbon emission, ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ refers to the anomalies of XCO₂ measured by the OCO-2 satellite, and the constant term k is added to the logarithmic function to avoid the situation where ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ is less than 0.

The local adaptive model is a critical step in integrating satellite-based XCO₂ observations with EDGAR carbon emission data. This approach assumes that the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ within each observation grid is primarily driven by emission activities occuring within that grid, and that a linear response exists between emissions and concentrations at the grid scale. It should be noted, however, that when the spatial extent of atmospheric transport exceeds the grid size (10 km), the potential associated errors may become non-negligible.

3.3. Spatial Redistribution of Carbon Emission Estimation

To generate a high-resolution carbon emission map, a spatial redistribution model was applied to refine the carbon emission estimates from a 10 km resolution (produced by the local adaptive model) down a finer resolution of 1 km.

Given that the spatial distribution of CO₂ concentration at the 1 km scale is inevitably influenced by atmospheric diffusion effects [29], the spatial redistribution model employs a weighting method based on high-resolution proxy datasets such as POIs, road networks, industrial site locations, and population distribution. This approach consists of two main steps: (1) sectoral allocation of total carbon emissions, and (2) spatial distribution of sector-specific emissions.

Since the local adaptive model outputs the total carbon emissions for each 10 km grid cell, these values need first be divided by sector proportion. For each 10 km grid cell j, the total carbon emissions are denoted as ${\mathrm{E}}_{10\mathrm{k}\mathrm{m}}\left[\mathrm{j}\right]$. Based on the sectoral proportion P_k provided by the CHRED database (see

Table 2. The spatial weighting methods for different emission sectors.

Sector-Specific	Spatial Proxy	Data Type	Sector Weights	Spatial Weights Methods
Commercial	POI	Point feature	0.0144	KDE and normalization
Industrial	Industry poi	Point feature	0.9366	KDE and normalization
Traffic	Roads data	Line feature	0.0254	KDE and normalization
Daily life	Population	Raster	0.0236	Normalization

), this total is allocated to different sectors k as follows:

$${\mathrm{E}}_{10\mathrm{k}\mathrm{m}}^{(\mathrm{k})}\left[\mathrm{j}\right]={\mathrm{E}}_{10\mathrm{k}\mathrm{m}}\left[\mathrm{j}\right]·{\mathrm{P}}_{\mathrm{k}}$$

where P_k represents the proportion of sector k (e.g., industry, transportation) in the total city-wide carbon emissions, and ${\mathrm{E}}_{10\mathrm{k}\mathrm{m}}^{(\mathrm{k})}\left[\mathrm{j}\right]$ denotes the carbon emissions of sector k in grid j.

For the spatial distribution of sector-specific emissions, a method was developed using high-resolution proxy data to represent spatial patterns of emissions.

First, for each 10 km grid cell j, all nested 1 km grid cells $\mathrm{i}\in \mathrm{j}$ are assigned spatial weights for sector k based on proxy data. These weights ${\mathrm{W}}_{\mathrm{k}}[\mathrm{i}]$ are then normalized as follows:

$${\mathrm{w}}_{\mathrm{k}}\left[\mathrm{i}\right]={\displaystyle \frac{{\mathrm{W}}_{\mathrm{k}}[\mathrm{i}]}{{\sum }_{\mathrm{i}\in \mathrm{j}}{\mathrm{W}}_{\mathrm{k}}[\mathrm{i}]}}$$

where ${\mathrm{W}}_{\mathrm{k}}[\mathrm{i}]$ represents the spatial proxy value for the 1 km grid cell i under sector k, calculated using kernel density estimation (as shown in Table 2); and ${\mathrm{w}}_{\mathrm{k}}\left[\mathrm{i}\right]$ is the normalized spatial weight, reflecting the relative contribution of the 1 km grid cell within its corresponding 10 km grid.

Next, the sector-specific emissions of the 10 km grid cell are allocated to its internal 1 km grids according to the spatial weights, using the following formula:

$${\mathrm{E}}_{1\mathrm{k}\mathrm{m}}^{(\mathrm{k})}\left[\mathrm{i}\right]={\mathrm{E}}_{10\mathrm{k}\mathrm{m}}\left[\mathrm{j}\right]·{\mathrm{w}}_{\mathrm{k}}\left[\mathrm{i}\right]$$

where $\mathrm{i}\in \mathrm{j}$ indicates that the 1 km grid cell i is nested within the 10 km grid cell j.

Finally, the total carbon emissions for each 1 km grid cell are obtained by aggregating emissions across all sectors:

$${\mathrm{E}}_{1\mathrm{k}\mathrm{m}}\left[\mathrm{i}\right]=\sum _{\mathrm{k}}{\mathrm{E}}_{1\mathrm{k}\mathrm{m}}^{(\mathrm{k})}\left[\mathrm{i}\right]$$

All variables are consistent with the definitions given above.

The spatial weighting methods for different emission sectors are shown in Table 2. The kernel density estimation (KDE) method was applied to POI, industry POI, and road network data, while population data could be used directly. All datasets were subsequently normalized. The sector weights were derived based on the ratio of carbon emissions of each department provided by the China High Resolution Emission Database (download: https://www.cityghg.com/).

3.4. Accuracy Assessment of the Model

The carbon emission estimation followed distinct methodological approaches at different scales. First, 10 km resolution estimates employ a statistics-fused local adaptive regression method; then, 1 km resolution estimates are derived through spatial proxy-based allocation. Given these fundamental methodological differences, accuracy assessments were performed separately for local adaptive model and spatial proxy allocation model.

The carbon emission estimates from the local adaptive model were compared with the EDGAR data. The relative error (RE) was calculated for each grid cell across different months using the following formula:

$$\mathrm{R}\mathrm{E}=({\mathrm{E}}_{\mathrm{L}\mathrm{A}}-{\mathrm{E}}_{\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{a}\mathrm{r}})/{\mathrm{E}}_{\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{a}\mathrm{r}}$$

where ${\mathrm{E}}_{\mathrm{L}\mathrm{A}}$ represents the estimation from local adaptive model, ${\mathrm{E}}_{\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{a}\mathrm{r}}$ represents the value of EDGAR, and RE is between −1 and 1.

The carbon emission estimates based on the spatial redistribution model were compared with ODIAC data. The Pearson correlation coefficient was calculated using the following formula:

$$\mathsf{\rho }={\displaystyle \frac{\mathrm{c}\mathrm{o}\mathrm{v}({\mathrm{E}}_{\mathrm{e}\mathrm{s}\mathrm{t}},{\mathrm{E}}_{\mathrm{o}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{c}})}{{\mathsf{\sigma }}_{\mathrm{e}\mathrm{s}\mathrm{t}}{\mathsf{\sigma }}_{\mathrm{o}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{c}}}}$$

where ${\mathrm{E}}_{\mathrm{e}\mathrm{s}\mathrm{t}}$ represents the estimation from spatial redistribution model, and ${\mathrm{E}}_{\mathrm{o}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{c}}$ represents the emissions of ODIAC, ${\mathsf{\sigma }}_{\mathrm{e}\mathrm{s}\mathrm{t}}$ and ${\mathsf{\sigma }}_{\mathrm{o}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{c}}$ represent the variances of the two respectively.

4. Results

4.1. The Trend of Background Concentration and Spatial Distribution of ${∆XCO}_2$

By overlaying spatial proxy data, we identified areas in Urumqi that are minimally impacted by human activities and calculated the average values in these regions to determine the background concentration of XCO₂. As shown in Figure 3, the left panel illustrates the intensity of human activity in Urumqi, while the right panel presented the corresponding average atmospheric XCO₂ concentrations in 2020.

It can be observed that human activities were most intense in the central part of Urumqi, followed by the southern area, whereas the northern part was less affected. The spatial distribution of the average XCO₂ concentration in 2020 shows a similar pattern, with higher values in the city center and lower values in surrounding areas, ranging from 411.51 ppm to 412.69 ppm. In areas with minimal human influence, the atmospheric XCO₂ concentration is approximately 412.07 ppm.

Figure 4 illustrates the trend of background concentration changes in Urumqi throughout 2020. As shown, the XCO₂ concentration in Urumqi ranged from approximately 408 ppm to 415 ppm during this period. From January to April (winter and spring), the XCO₂ concentration increased slightly each month, rising from 412.98 ppm to 414.99 ppm. Between April and August (spring and summer), it showed a significant month-by-month decrease, reaching a low of 408.15 ppm. However, from August to December (autumn and winter), the concentration trended upward again, closing the year at approximately 414.67 ppm in December.

By analyzing the changing trend of XCO₂ concentration values, we can find an obvious seasonal variation pattern of “high in spring and winter, low in summer and autumn”. In summer and autumn, the atmospheric XCO₂ concentration is at a low point due to vegetation growth and strong photosynthesis, while the atmospheric XCO₂ concentration increases in spring and winter, on the one hand due to the weakening of vegetation growth, and on the other hand due to human heating needs. The XCO₂ concentration decreases from April until August (spring and summer), i.e., the period of vegetation growth, and XCO₂ concentrations increase from September until December (autumn) and from January until April (winter) due to reduced vegetation photosynthesis.

The XCO₂ spatial distribution map of each month is subjected to raster difference calculation, that is, the XCO₂ spatial distribution is subtracted from the background concentration value of the corresponding month to obtain the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ of each month. It can analyze the spatial distribution of atmospheric CO₂ concentration after excluding seasonal changes, which to a certain extent reflects the changing pattern of carbon emissions under the assumption that it is not affected by atmospheric diffusion. Figure 4 shows the spatial distribution of ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ from January to December 2020.

As shown in Figure 5, the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ ranged from −1.58 ppm to 1.13 ppm in 2020. In spring, autumn, and winter (January to May and October to December), ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ in Urumqi showed a spatial distribution of “positive in the north and negative in the south”. According to the statistical data of wind speed and direction in Urumqi from 1987 to 2017 for a total of 30 years by the ground meteorological station No. 51463 (43.7892N, 87.6458E, 935.0 m), there is more calm weather in winter, low average wind speed, and more northeast and southwest wind directions; the wind direction in spring and autumn is not concentrated, and the wind speed is not high. Therefore, the XCO₂ concentration is not greatly affected by wind speed and direction in these months, and the spatial distribution of ${∆\mathrm{XCO}}_2$ can reflect the spatial distribution of carbon emissions to a certain extent. In other words, the carbon emission intensity in the city center and the north is higher than that in the south of the city. The results showed that the dominant wind direction and the wind direction with the maximum wind speed in Urumqi in summer are both northwest winds, with high average wind speeds (about 3.6 m/s), which have a more obvious effect on the horizontal transmission of atmospheric CO₂. Therefore, in the summer (June to September), carbon emissions from the urban center of Urumqi diffuse southward under the influence of wind speed and direction, and the atmospheric CO₂ concentration in the southern part of the city increases above its background concentration, that is, the value of ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ is positive.

Notably, atmospheric transport exerts a considerable influence during the summer months (June to September). Consequently, when conducting regression modeling on fixed grids using ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ data from this period, it is essential to account for associated uncertainties. A detailed discussion of this issue is provided in Section 5.1.

4.2. Cluster Analysis of ${∆XCO}_2$ and EDGAR Carbon Emission Time Series

The spatial and temporal distribution analysis of ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ further verified that it has a certain correlation with carbon emissions. On this basis, the K-mean method was used to cluster the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ time series and the carbon emission time series in Urumqi, and a relationship model between the two was constructed. According to the clustering results, the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ time series and the carbon emission time series basically presented three types (Cluster A, B, and C), as shown in Figure 6.

As shown in Figure 6, the relationship between the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ time series and the carbon emission time series is completely different in Cluster A and the other two categories (B and C). In Cluster A, from January to August, ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ gradually increased from negative to positive, that is, the role of carbon sink capacity gradually weakened, from −0.7 ppm to −0.2 ppm, and then increased to 0.7 ppm; from August to December, ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ gradually decreased from positive to negative (about −0.6 ppm); among them, ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ from June to September was positive, which is closely related to the wind speed and direction mentioned in Section 4.1. In Cluster B and Cluster C, the relationship between the ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ time series and the carbon emission time series is very similar; the only difference is that they correspond to different orders of magnitude of carbon emissions and the size of ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$. The carbon emission level of Cluster B is 10 × 10⁴ tons, while the carbon emission level of Cluster C is 1 × 10⁴ tons. The two show the same overall trend of change, that is, as carbon emissions decrease, it can be observed that ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ is also decreasing, and at the same time, an increase in carbon emissions will also lead to an increase in ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$.

We have drawn a spatial distribution map of different cluster types, as shown in Figure 7. As shown in Figure 7, Cluster A is distributed in the southern part of Urumqi, Cluster B is the urban center and industrial cluster area of Urumqi, and Cluster C is in the periphery of the urban center and the northern part of the city. Among them, the central location, where Cluster B is located, has the most intensive and active human activities and developed industry, commerce, and transportation; the surrounding areas and the northern part of the city, namely Cluster C, are also within the scope of human activity, but the human activities in the city center will be reduced; the southern part of the city is higher in terrain, with sparse vegetation such as grassland and shrubs, and bare land. While vegetation coverage in this area is limited and consists mostly of low-biomass types, the reduced anthropogenic activity may contribute more significantly to the lower observed CO₂ concentrations. The southern part of the city is surrounded by mountains on both sides and a valley in the middle, so there is valley wind all year round. Under the influence of valley wind, carbon emissions from far away are transmitted to the local area, diluting the local carbon sink effect [30]. From January to August, as the average wind speed increases, the local carbon sink is diluted more and more, resulting in a change that is completely opposite to the local carbon emissions.

4.3. Carbon Emissions Spatial Distribution Map in Urumqi

The local adaptive model was utilized to integrate the ${∆\mathrm{XCO}}_2$ and carbon emission from EDGAR, resulting in the distribution of carbon emissions with 10 km spatial resolution. As shown in Figure 8, the carbon emissions with 10 km resolution distribution map from January to December 2020 has been drawn. The spatial distribution of carbon emissions in Urumqi throughout the year presented a pattern of high emissions in the city center and low emissions in the surrounding areas. The monthly carbon emissions of a single grid (10 km × 10 km) in the high-value area exceed 357,000 tons, and the carbon emissions within individual grids are higher than 665,000 tons. The monthly carbon emissions of a single grid in the low-value area are less than 23,000 tons.

Furthermore, the carbon emissions estimation with 1 km resolution was generated under the spatial redistribution model in Section 3.3. As shown in Figure 9, the estimated carbon emission distribution at a spatial resolution of 10 km and the estimated spatial carbon emission distribution at 1 km in January 2020 are plotted. The grid cell with the most carbon emissions in Figure 9a is about 1.19 mt, while the grid with the most carbon emissions in Figure 9b is about 44.65 kt after further spatial allocation.

As shown in Figure 9, in the spatial redistribution process based on the 10 km carbon emission estimation, the spatial proxy data played a more obvious role. The clustered hot spots of the proxy data are still clustered in the redistribution of carbon emissions, and some carbon emission point sources are found in the 1 km resolution map.

5. Discussions

The carbon emission was estimated successively by the local adaptive model and spatial redistribution model. Since the error sources were different, the 10 km spatial resolution carbon emission estimation and 1 km spatial resolution carbon emission estimation were validated separately, and discussed the accuracy and limitations.

5.1. Validation of Estimates with EDGAR

The local adaptive model assumes that the observed ${∆\mathrm{XCO}}_2$ within a given grid is primarily driven by emission activities inside that grid, and that emissions and concentrations exhibit a linear response relationship at the grid scale. However, when the spatial extent of atmospheric transport exceeds the grid size (10 km), the resulting potential errors cannot be ignored. Therefore, this study focuses on evaluating the estimation accuracy of the local adaptive model in comparison with the EDGAR carbon emission data, and analyzes monthly estimation errors to assess the influence of atmospheric dispersion on the results.

The total carbon emissions for Urumqi in 2020 based on the EDGAR dataset was calculated, which amounts to approximately 58.26718 million tones. In comparison, the estimated total emissions using the local adaptive model were about 58.26714 million tones, indicating a very close agreement between the two datasets.

The relative errors of the carbon emission distribution by the local adaptive model (10 km × 10 km) and from EDGAR in different months were counted, so as to analyze and measure the deviation ratio between the estimated value and the EDGAR result. Figure 10 plotted the relative error distribution of the two in different months, and we assume an acceptable error range (plus or minus 0.1). The errors of the estimated values from January to December are basically within the acceptable range, and the maximum deviation occurs in January, about 1.37.

From the results in Figure 10, it can be seen that the relative error distribution between ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ and EDGAR were concentrated between −0.1 and 0.1, and the overall distribution is positive and long-tailed due to a small number of extreme values. In other words, the estimated value of carbon emissions in this study are generally close to those of EDGAR; the model can integrate ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ from satellite observations and EDGAR from statistical data and has a good effect in relative error assessment. It is worth noting that the emergence of the positive long-tail distribution further confirmed the reliability of our fusion results, where some grid-level estimates exceed those in the EDGAR dataset. This may be attributed to the fact that the version of EDGAR used in this study does not include carbon emissions from biomass burning. In addition, previous studies have suggested that EDGAR may underestimate emissions in industrialized cities [31,32].

To further analyze the spatial distribution of estimation errors, Figure 11 presents the spatial distribution of the relative error between the estimates from the local adaptive model and the EDGAR carbon emissions from January to December 2020. As shown in the figure, the relative errors for most months were close to zero, which was consistent with the results in Figure 10.

In addition, the spatial distribution of relative errors between the local adaptive model and EDGAR indicates a seasonal variation pattern. From April to September (spring and summer), the wind activity and atmospheric convection are generally stronger and the relative errors in high-emission in central areas were predominantly positive. This result was consistent with a study that used the Gaussian plume diffusion model to compare the relationship between satellite XCO₂ concentration and carbon emissions, suggesting that in regions prone to atmospheric diffusion, EDGAR data may be underestimated [33]. In contrast, during the other months, the relative errors in high-emission areas were mostly negative.

It can be seen from this that the local adaptive model may to some extent correct the underestimation of the EDGAR data. However, the spatial heterogeneity of relative errors and their seasonal variations indicate that atmospheric diffusion remains a non-negligible factor [34]. Therefore, future studies should incorporate diffusion models to further improve the accuracy and physical reliability of the estimates.

5.2. Validation of Estimates with ODIAC

In order to objectively evaluate the estimated data of this study, we downloaded the ODIAC data for 2020 and compared it with the data estimated in this study. The spatial resolution of both datasets is 1 km × 1 km. As shown in Figure 12, (a) is the carbon emission distribution of the central urban area of Urumqi in 2020 (derived from this study), and (b) is the carbon emission distribution of the central urban area of Urumqi in 2020 (derived from ODIAC data).

Figure 12 revealed noticeable discrepancies in the spatial distribution of carbon emissions in the urban center between the two datasets. The correlation coefficient was approximately 0.192, suggesting a weak spatial association.

Figure 12 fixes the position and scale. As shown in Figure 12a, the carbon emission estimation results of this study mainly present four high carbon emission areas, which are located at the positions marked as ➀–➃ in the figure, scattered around the city rather than clustered. As can be seen from Figure 12b, ODIAC’s carbon emissions are clustered, which is a different pattern from this study. The ODIAC carbon emission distribution was derived from nighttime light data and the distribution of large power plants. Although this is a recognized method for estimating carbon emissions, it ignores different types of CO₂ emissions, such as commercial carbon emissions, industrial carbon emissions, and residential carbon emissions. Therefore, in the ODIAC carbon emission distribution map, there was a relatively smooth pattern of gradual decrease from the city center to the periphery of the city, and large power plants in the city were individually assigned values far greater than the carbon emissions of the city center. The method proposed in this study further used POI, factories, population density, and road distribution data after integrating carbon satellite observations and EDGAR statistical values, and can secondarily allocate carbon emissions according to different types of carbon emission sources.

In order to make a more objective evaluation of the two, the rationality of the above carbon emission distribution is discussed with the current urban land use map (Figure 13) as a medium. Figure 13 is the current land use map of the central urban area of Urumqi in 2018 [35]. Due to the difficulty in obtaining the current land use map in 2020, we only use Figure 13 as a reference and assume that the city is less likely to undergo a significant land use change within two years.

As shown in Figure 13, industrial land in Urumqi is distributed around the city. The distribution of the three high carbon emission sources in Figure 13 corresponds exactly to the distribution of the three industrial zones in Figure 12(➀➁➃). The high carbon emission source ➂ in Figure 12a corresponds to the urban commercial center ➂ in Figure 13. These findings do not have a similar correspondence in Figure 12b. Due to the inherent characteristics of night light distribution and the separate estimation of large power plant point sources in ODIAC mapping, the carbon emission distribution in Figure 12b shows a smaller variance within the city. Therefore, even if its spatial resolution has reached 1 km, the availability of meaningful data at the urban scale is still limited. Therefore, it is worth affirming that the carbon emission distribution data with a spatial resolution of 1 km × 1 km generated by the method proposed in this study at the urban scale can reflect the characteristics of the spatial distribution of urban carbon emissions better than the ODIAC data with the same resolution, thereby providing a data basis for policymakers to make scientific carbon reduction plans.

5.3. Methodological Limitations and Future Perspective

This study integrated the local adaptive model with a spatial redistribution approach to establish the response relationship between XCO₂ concentration variations and anthropogenic emissions within specific grid cells. The spatial redistribution method further refined the estimated emissions to a 1 km spatial resolution. Given the considerable inconsistencies among current carbon emission databases [36,37], this study provided empirical evidence to support the potential of carbon satellites in quantitatively monitoring anthropogenic emissions. It also offered a promising approach for independent verification of carbon emission inventories and provided new insights into the fine-scale monitoring of urban-scale carbon emissions.

However, this method still presented certain limitations under the current research framework.

The estimation relied primarily on the EDGAR global carbon emission database as the baseline. The results suggested that the local adaptive model may partially correct the underestimation inherent in EDGAR. Nevertheless, carbon emission datasets like EDGAR were subject to considerable regional uncertainties, and related errors require further verification [38].

As discussed in Section 5.1 on the validation of the local adaptive model, the impact of atmospheric diffusion cannot be ignored. The current model established a statistical relationship between ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ and anthropogenic emissions based solely on regression, without incorporating physical driving mechanisms such as atmospheric transport [33,34].

In addition, the calculation of background concentrations and ${∆\mathrm{X}\mathrm{C}\mathrm{O}}_2$ was a critical step in reducing the uncertainty of emission estimates. This process may require adjustments based on regional factors such as vegetation carbon sinks, the intensity of human activities, and topographic features [39,40].

6. Conclusions

This study proposed a framework that integrated the local adaptive model with a spatial redistribution method to estimate urban carbon emissions from OCO-2 observations. The estimated emissions were validated with EDGAR and ODIAC datasets, and the results indicated the following: (1) The total estimated carbon emissions in Urumqi in 2020 differed from EDGAR data by only 20 tons, with monthly relative errors generally within ±10%, indicating high accuracy. (2) The seasonal distribution of relative errors showed positive deviations in high-emission areas during spring and summer, and negative deviations during autumn and winter. (3) The spatial distribution of emissions after redistribution is not entirely consistent with ODIAC, with a Pearson r = 0.192. (4) A comparative analysis with current urban land use maps confirms that the emission patterns derived in this study were reasonable.

This approach may identify potential biases and inconsistencies in city-scale emission inventories derived from bottom-up methods. However, several limitations remain, including the influence of atmospheric diffusion on estimation results and the model’s dependence on data sources such as EDGAR. Future work will extend the application of this method to additional cities and incorporate atmospheric transport models to better quantify the errors and uncertainties in carbon emission estimates.