Detecting the Response of Column Carbon Dioxide Concentration to Anthropogenic Emissions Using the OCO Series Satellites

Highlights

What are the main findings?

• A novel LISA-based framework is proposed to quantify local anthropogenic CO₂ increments (dXCO₂) within urban agglomerations using a seamless satellite XCO₂ dataset.
• Significant spatial heterogeneity in dXCO₂ is observed across China’s three major urban agglomerations, with the northern heavy industrial region (BTH: 3.46 ppm) exhibiting much higher values than the southern coastal economic zones (YRD: 1.3 ppm; PRD: 0.08 ppm) during 2015–2019.
• Four dominant modes governing XCO₂ anomalies are identified via EOF analysis, revealing multi-scale influences from anthropogenic–natural trade-offs, the East Asian summer monsoon, local emission events, and geographical background.
• Differential responses to COVID-19 lockdowns are captured, with dXCO₂ decreasing in BTH and YRD but increasing in PRD, highlighting the modulating role of regional industrial structures.

What are the implications of the main findings?

• The LISA-based method provides a new technical pathway for regionally differentiated emission reduction assessments by enabling the extraction of anthropogenic signals from background variability.
• The observed spatial–temporal patterns and the influence of industrial structures on emission resilience offer crucial insights for designing targeted and effective climate mitigation policies in complex urban agglomerations.
• The successful application of the seamless XCO₂ dataset demonstrates its strong potential for monitoring long-term CO₂ changes and verifying emission reduction efforts from space.

Abstract

Quantifying anthropogenic CO₂ increments is vital for assessing emission reductions. Using a seamless XCO₂ dataset over China reconstructed from OCO-2/3 satellite retrievals and machine learning, combined with EOF decomposition and LISA analysis, this study investigates XCO₂ anomalies and local anthropogenic increments (dXCO₂) at national and urban agglomeration scales. Nationally, XCO₂ anomalies exhibit a “southeast positive, northwest negative” spatial pattern aligning with human activities and a “winter high, summer low” seasonal cycle. EOF analysis reveals four dominant modes: anthropogenic–natural trade-offs, East Asian summer monsoon modulation, local emissions, and baseline context. At the regional scale, multi-year mean dXCO₂ (2015–2019) in Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD), and Pearl River Delta (PRD) are 3.46 ± 0.45 ppm, 1.30 ± 0.36 ppm, and 0.08 ± 0.14 ppm, respectively, showing higher values in northern heavy industrial zones. During the 2020–2022 pandemic, dXCO₂ decreased in BTH (2.28 ± 0.73 ppm) and YRD (1.16 ± 0.43 ppm) but increased in PRD (0.28 ± 0.27 ppm). Compared to pre-pandemic levels, lockdowns saw dXCO₂ decrease slightly in YRD while increasing in BTH and PRD, reflecting differential responses of regional industrial structures. This study demonstrates the potential of seamless XCO₂ data for monitoring anthropogenic enhancement signals, and the proposed LISA-based method offers new support for regionally differentiated emission reduction assessments.

1. Introduction

Since the Industrial Revolution, global greenhouse gas emissions have surged, and the consequent climate change has been scientifically substantiated [1]. Among these, CO₂, as the primary greenhouse gas emitted from fossil fuel combustion, contributes over 60% to total global radiative forcing [2]. Human activities are the main source of CO₂ emissions. Cities, as concentrated areas of human activity, host more than half of the global population and the vast majority of economic activities, contributing over 70% of CO₂ emissions despite occupying only a small fraction of the Earth’s land surface [3]. The Paris Agreement explicitly sets the goal of limiting global average warming to well below 2 °C and pursuing efforts to limit it to 1.5 °C [4], requiring countries to achieve net-zero emissions by mid-century. To meet this goal, accurately quantifying the increment in atmospheric CO₂ concentration caused by human activities, assessing the effectiveness of national emission reduction efforts, and identifying key emission areas have become critical scientific supports for climate governance.

Traditional “bottom-up” observation methods based on ground-based monitoring networks are limited by sparse station distribution, making it difficult to cover emission hotspot areas and resolve the spatial heterogeneity of anthropogenic emissions. The emergence of carbon satellites has provided a new pathway for large-scale, high-precision observation of atmospheric CO₂, establishing them as crucial tools for current global carbon cycle research and carbon emission monitoring. For instance, Hakkarainen et al. utilized OCO-2 data to achieve, for the first time, direct spatial observation of localized CO₂ emission areas [5]; Schwandner et al. confirmed the capability of spaceborne platforms to detect localized CO₂ emission sources [6]. “Top-down” inversion methods based on carbon satellite observation data have become a key technical means for independently assessing and verifying national and regional emission reports.

By differencing observed XCO₂ against background concentrations, carbon satellites can effectively identify and quantify XCO₂ enhancement signals resulting from both natural processes and human activities. This method has become an important paradigm in current carbon cycle research. Sheng et al., based on OCO-2 data over the Yangtze River Delta urban agglomeration in China, found that XCO₂ around major cities and industrial activity intensive areas increased by an average of 1.36~4.41 ppm [7]; Ye et al., focusing on the Xiamen–Zhangzhou–Quanzhou metropolitan area and utilizing high-resolution snapshot data from OCO-3, observed XCO₂ enhancements ranging from 0.70 ± 0.53 ppm to 2.29 ± 1.16 ppm, and further constrained anthropogenic CO₂ emissions in the Xiamen area using Bayesian inversion methods [8]; Ji et al., by analyzing 473 global OCO-2 observation cases from 2014 to 2023, successfully retrieved point-source scale CO₂ emissions, finding a global average emission rate of 6166.52 t CO₂/h, with case studies in California, Riyadh, and Xinjiang further validating the effectiveness of satellite differencing methods in capturing localized emission plumes [9]. At the urban scale, Xu et al. utilized OCO-3 XCO₂ observations combined with Gaofen-5 and TROPOMI NO₂ data to identify enhancements from 69 global emission sources, showing that anthropogenic emission signals in regions such as East Asia, Europe, the Middle East, and North America were effectively captured, with approximately 45% of estimated emissions lower than inventory reports, reflecting the potential of satellite observations for independent verification [10]. These studies indicate that by differencing satellite-observed XCO₂ from appropriate background values, the resulting increments can effectively reflect contributions from different sources, providing a key technical means for distinguishing between natural and anthropogenic signals.

However, constrained by satellite observation conditions and scanning modes, most existing studies have been limited to point-source scales [11]. Consequently, effectively identifying anthropogenic enhancement signals at the urban agglomeration scale from continuous XCO₂ fields remains a pressing challenge. To address this issue, this study leverages a high-resolution, spatially and temporally continuous XCO₂ dataset over China—constructed from OCO-2 and OCO-3 satellite retrievals—to systematically investigate the spatiotemporal evolution of XCO₂ increments induced by human activities at both national and intra-urban agglomeration scales. Specifically, spatial overlay analysis is first employed to examine the spatial correspondence between areas of elevated XCO₂ concentrations and high carbon emission intensities, thereby assessing the capability of the concentration field to indicate ground-level emission sources across different regions. Subsequently, empirical orthogonal function (EOF) decomposition is applied to extract the dominant spatiotemporal structures of the XCO₂ anomaly field and to identify the principal modes of variability and their temporal evolution. Building upon this foundation, a technical framework integrating high-resolution XCO₂ spatial distributions with spatial autocorrelation analysis is proposed, with a focus on three major urban agglomerations—the Beijing–Tianjin–Hebei region, the Yangtze River Delta, and the Pearl River Delta. This framework aims to isolate local anthropogenic CO₂ increments within urban agglomerations and to characterize their spatial heterogeneity, thereby providing a scientific basis for the formulation of regionally differentiated emission reduction policies.

2. Materials and Methods

2.1. Data

2.1.1. XCO₂ Dataset

The TAF-XCO₂ (Temporally Adaptive Forest) [12] dataset employed in this study spans the China region from January 2015 to December 2022, with a spatial resolution of 0.01° × 0.01° and a monthly temporal resolution. This dataset is a seamless reconstruction product generated using the Temporally Adaptive Forest method based on XCO₂ retrievals acquired by the OCO-2 satellite [13,14] (January 2015–December 2022) and the OCO-3 satellite [15] (September 2019–December 2022); comprehensive methodological details have been documented in a dedicated data descriptor publication. Specifically, the OCO-2 XCO₂ retrievals were obtained from the OCO₂_L2_Lite_FP product (v11.1r), and the OCO-3 XCO₂ retrievals from the OCO3_L2_Lite_LP product (v10.4r), both of which were provided by the Goddard Earth Sciences Data and Information Services Center (GES DISC, https://disc.gsfc.nasa.gov/). Previous investigations have indicated that the XCO₂ estimates from the two satellites differ by approximately 0.2 ppm over land and 0.3 ppm over ocean [16]. Despite the non-overlapping observational periods of the two satellites, their data were jointly utilized in this study to enhance the completeness and continuity of spatiotemporal coverage.

The TAF-XCO₂ dataset was constructed based on the Temporally Adaptive Forest (TAF) model [12]. The overall workflow is illustrated in Figure 1. First, a unified input layer was established by resampling all environmental variables to an identical spatial resolution and coordinate system. Subsequently, feature-layer optimization was performed: for categorical variables, an annually dynamic one-hot encoding strategy was adopted to ensure optimal alignment between feature representation and the data distribution of the corresponding year; for continuous variables, standardization was applied to eliminate the influence of disparate dimensional units. On this basis, a grid search coupled with three-fold cross-validation was employed to optimize the model hyperparameters on an annual basis, including the number of decision trees, maximum tree depth, and the minimum number of samples required at a leaf node. The training dataset was randomly partitioned into three subsets, with two subsets serving as the training set and the remaining subset as the validation set. The optimal hyperparameter combination for each year was determined by minimizing the root mean square error (RMSE). The annually optimized hyperparameter results are presented in Table A1. Using the optimized hyperparameters, independent regression models were constructed for each year and trained on the respective annual training sets to learn the complex nonlinear mapping relationships between multi-source environmental variables, including vegetation indices (e.g., NDVI, EVI, GPP), meteorological elements (Temperature, Precipitation, Wind speed, Potential evapotranspiration, and Surface pressure), emission-related variables (Fossil fuel emission, NO₂), and topographic factors (digital elevation model), and XCO₂. After model training, the independent variable data for all twelve months of a given year were fed monthly into the corresponding annual TAF model to generate monthly XCO₂ prediction raster images. Ultimately, through the sequential output of independently trained annual models, a monthly seamless XCO₂ dataset for China at a spatial resolution of 0.01° was produced for the period 2015–2022.

Figure 1. Workflow for reconstructing XCO₂ distribution using satellite XCO₂ retrieval data.

Independent validation against ground-based station observations demonstrates that the TAF-XCO₂ data exhibit strong consistency with TCCON measurements: at the Hefei station, the R² value reaches 0.89 with an RMSE of 1.387 ppm, while at the Xianghe station, the R² is 0.737 with an RMSE of 2.026 ppm.

The CAMS XCO₂ data (Copernicus Atmosphere Monitoring Service) is a global gridded atmospheric CO₂ column concentration reanalysis product generated by assimilating multi-source satellite and ground-based observations combined with atmospheric transport models. It offers advantages such as spatiotemporal continuity and scientific reliability, providing crucial data support for carbon cycle research, satellite validation, and emission estimation. In this study, monthly composites of CAMS XCO₂ data are used for comparative analysis of spatiotemporal variability with TAF XCO₂. This dataset has a spatial resolution of 0.75° × 0.75°, covering the period from January 2015 to December 2020, and is sourced from the European Centre for Medium-Range Weather Forecasts (https://ads.atmosphere.copernicus.eu/, accessed on 28 April 2026).

All datasets utilized in this study are summarized in Table 1.

Table 1. Data Sources Used for XCO₂ Reconstruction.

Data Type	Parameter	Spatial Resolution	Temporal Resolution	Source	Time Range
Orbiting Carbon Observatory-2 data	OCO-2_L2_Lite_FP (v11.1r)	2.25 km × 1.29 km	Daily	NASA GES DISC (https://disc.gsfc.nasa.gov) (accessed on 1 July 2025)	January 2015–December 2022
Orbiting Carbon Observatory-3 data	OCO3_L2_Lite_LP (v10.4r)	2.25 km × 1.29 km	Daily		September 2019–December 2022
Vegetation data	NDVI	500 m	16 days	MOD13A2 (https://modis.gsfc.nasa.gov/) (accessed on 4 July 2025)	January 2015–December 2022
	EVI	500 m	16 days
	GPP	500 m	16 days
Meteorological data	Temperature	About 1 km	Monthly	National Qinghai–Tibet Data Center (https://data.tpdc.ac.cn/) (accessed on 8 July 2025)	January 2015–December 2022
	Precipitation	About 1 km	Monthly
	Potential evapotranspiration	About 1 km	Monthly
	Wind velocity	About 1 km	1 h	ERA5 monthly averaged data (https://cds.climate.copernicus.eu/datasets) (accessed on 7 July 2025)	January 2015–December 2022
	Surface pressure	About 1 km	Monthly
Emission data	Fossil fuel emission	0.01° × 0.01°	Monthly	ODIAC2023 (https://db.cger.nies.go.jp/dataset/ODIAC/) (accessed on 9 July 2025)	January 2015–December 2022
	NO2	0.01°× 0.01°	Monthly	National Qinghai–Tibet Data Center (https://data.tpdc.ac.cn/) (accessed on 8 July 2025)	January 2015–December 2022
Topography data	DEM	90 m		SRTM V3 (https://www.gscloud.cn/)(accessed on 8 July 2025)	November 2013
Validation data	TCCON XCO2	Point	About 2 min	TCCON Observation Network (https://tccondata.org) (accessed on 15 July 2025)	November 2015–December 2022
	CAMS XCO2	0.75°× 0.75°	3 h	ERA5 monthly averaged data (https://ads.atmosphere.copernicus.eu/)(accessed on 15 July 2025)	January 2015–December 2020

2.1.2. Carbon Emission Data

EDGAR (Emission Database for Global Atmospheric Research) is an authoritative global anthropogenic greenhouse gas emissions dataset developed and maintained by the Joint Research Centre (JRC) of the European Commission [17]. This database (https://aqmx.org) provides long time-series emission data from 1970 to the present. Its latest version integrates collaborative results from the JRC and the International Energy Agency (IEA) [18], covering a full range of major greenhouse gases including carbon dioxide (CO₂), methane (CH₄), nitrous oxide (N₂O), and fluorinated gases, with a spatial resolution of 0.1° × 0.1°. This study utilizes the annual carbon emission data from this dataset for the period 2015–2022.

2.2. Methods

To systematically resolve the spatial structure of XCO₂ anomalies, this study employs three complementary analytical methods in sequence. First, spatial overlay analysis is conducted to evaluate the spatial consistency between concentration hotspots and emission sources, thereby clarifying the applicable premise of anthropogenic increment detection. Next, empirical orthogonal function (EOF) analysis is applied to isolate the dominant modes of variability across different scales. Finally, the local indicator of spatial association (LISA) is utilized to extract localized increment signals attributable to anthropogenic emissions.

2.2.1. Spatial Overlay Analysis

This study aims to effectively extract the local enhancement signals attributable to anthropogenic emissions from the continuously distributed XCO₂ concentration field. It should be noted that satellite-observed XCO₂ column concentrations inherently result from the combined effects of surface emissions, atmospheric transport processes, and natural sources and sinks. Consequently, in the absence of fundamental spatial consistency between regions of elevated XCO₂ concentrations and hot pixels of anthropogenic carbon emissions, subsequent analyses that seek to identify anthropogenic enhancement based on the concentration field would lack the requisite prerequisite foundation. To quantitatively assess the degree of spatial correspondence between the two, this study employs a spatial overlay analysis method. Specifically, this approach calculates the proportion of the overlapping area between pixels with high XCO₂ concentrations and those with high carbon emissions relative to the total area of high-emission pixels, thereby establishing a quantitative metric that characterizes the spatial consistency between emissions and observed concentrations. The calculation formula is as follows:

$$Overlap Ratio={\displaystyle \frac{{XCO}_{2hot pixels}\cap Emissio{n}_{hot pixels}}{Emissio{n}_{total pixels}}}\times 100\%$$

(1)

The $Overlap Ratio$ (OR) effectively identifies areas potentially influenced by anthropogenic emissions contributing to the increase in regional atmospheric CO₂ concentration.

2.2.2. Empirical Orthogonal Function Analysis

The spatiotemporal variability of XCO₂ is subject to the compounded influences of monsoonal transport, vegetation phenology, interannual climate oscillations, and anthropogenic emission fluctuations. Analyzing these individual processes is therefore essential to mitigate the background interference imposed by natural variability. To this end, the empirical orthogonal function (EOF) method is introduced. Empirical Orthogonal Function analysis is a statistical method for parsing the spatiotemporal variability of large, multi-dimensional datasets and is widely applied in various geographical fields such as atmospheric science and oceanography [19,20,21]. As a form of principal component analysis, the EOF method decomposes multi-dimensional data into a set of orthogonal spatial patterns (EOFs) and their corresponding temporal coefficients (PCs), ordered by their variance contribution. This ensures that the leading modes explain the majority of the variance in the original dataset [22].

Let the original data matrix be $X$. First, its spatial covariance matrix $R_s$ is calculated:

$$R_s={\displaystyle \frac{1}{m}}{X}^{T}X$$

(2)

where $R$ denotes the covariance matrix, the subscript $s$ indicates the spatial domain, the superscript $T$ denotes matrix transpose, and $m$ is the number of spatial samples. Eigendecomposition is performed on the covariance matrix:

$$R_S{E}_S=L_S{E}_S$$

(3)

Here, $E$ is the eigenvector matrix, whose column vectors correspond to the temporal coefficients (PCs); $L$ is a diagonal matrix whose diagonal elements are the eigenvalues. Since the dimension of the covariance matrix equals the number of observations, each eigenvector represents a time series.

Projecting the spatial anomaly matrix $X$ onto the eigenvector matrix $E$ yields the corresponding spatial modes (EOFs):

$$F_s=X{E}_S$$

(4)

where each column of $F$ represents a spatial mode. The original data can be reconstructed from the spatial modes and temporal coefficients:

$$X=F_S{E}_S^T$$

(5)

The variance contribution rate of each EOF/PC mode is obtained by dividing its corresponding eigenvalue by the sum of all eigenvalues. Modes are ordered by the amount of variance they explain; the first mode explains the largest fraction of variance, the second mode explains the next largest, and so on. This ordering property enables the EOF method to effectively extract dominant spatiotemporal variability signals, rendering it well-suited for identifying the principal modes of variation in complex regions and thereby providing a crucial characterization of the background field for the subsequent incremental separation focused on the urban agglomeration scale.

2.2.3. Identifying Local Anthropogenic Emission Increments Based on LISA

Spatial autocorrelation is a statistical metric for measuring spatial dependence, which can be categorized into global autocorrelation and local autocorrelation based on spatial scale [23]. To identify the degree of autocorrelation between each spatial unit and its neighboring units, and to quantitatively characterize the spatial association differences across various regions, this study employs the Local Moran’s I (LISA) method [24,25]. For a spatial location $i$, the calculation formula for LISA is:

$$I_i=Z_i{{\displaystyle \sum }}_{j=1}^n{W}_{ij}{Z}_j$$

(6)

where $Z\left(I_i\right)$and$Z_j$are the standardized deviations of the observed values from the mean for spatial units $i$ and $j$, respectively; $n$ denotes the total number of pixels; $W_{ij}$ is the spatial weight matrix, which is row-standardized during model computation.

The test statistic for the significance of LISA is the $Z\left(I_i\right)$ score, calculated as follows:

$$Z\left(I_i\right)={\displaystyle \frac{I_i-E\left(I_i\right)}{\sqrt{VAR\left(I_i\right)}}}$$

(7)

where$I_i$ is the LISA value, $E\left(I_i\right)$ is its expected value, and $VAR\left(I_i\right)$ is its variance.

Under the condition that the $Z\left(I_i\right)$ score is significant, the LISA result can be interpreted as four types of spatial association [26]: ① When $I_i$ > 0 and $Z\left(I_i\right)$ > 0, it indicates that unit $i$ and its neighbors are both high values, representing a High–High cluster (HH); ② When $I_i$ > 0 and $Z\left(I_i\right)$ < 0, it indicates that unit $i$ and its neighbors are both low values, representing a Low–Low cluster (LL); ③ When$I_i$ < 0 and $Z\left(I_i\right)$ > 0, it indicates that unit $i$ is a high value while its neighbors are low values, representing a High–Low outlier (HL); ④ When $I_i$ < 0 and $Z\left(I_i\right)$ < 0, it indicates that unit $i$ is a low value while its neighbors are high values, representing a Low–High outlier (LH).

Ideally, CO₂ concentrations should be elevated in proximity to emission sources and decline with increasing distance, exhibiting a spatial distribution consistent with Tobler’s First Law of Geography. Traditional background subtraction methods typically rely on the mean concentration within a single latitudinal band as the background value. This approach introduces substantial biases in regions characterized by dense emission sources interspersed with surrounding natural carbon sinks. To overcome this limitation, the present study applies LISA to the analysis of XCO₂ data in order to identify spatial clustering patterns of elevated XCO₂ values. A significance threshold of p ≤ 0.05 was adopted for cluster classification. By objectively delineating the spatial extent of emission hotspots directly from the continuous field, and by accounting for underlying surface heterogeneity, the LISA approach enables a more robust isolation of the net contribution of local anthropogenic emissions to XCO₂ concentrations [7]. Since the reconstructed data is in GeoTIFF format, it needs to be converted to point features before analysis to calculate the Local Moran’s I index. Based on this, HH clusters are considered as local anthropogenic emission source areas, and the mean XCO₂ value within LL clusters is taken as the background concentration. The background value is extracted using a latitudinal zonal statistics method: first, the mean is calculated for LL points within each 0.01° latitude interval, and then a polynomial fit is applied to construct a background concentration curve across continuous latitudinal bands. The calculation formula for the local anthropogenic CO₂ increment (${\mathrm{dXCO}}_2$) is:

$${\mathrm{dXCO}}_2=XC{O}_2^{HH}-XC{O}_{2, BK}^{LL-mean}$$

(8)

3. Results

3.1. Spatiotemporal Characteristics of XCO₂ Across China

The grid-level differences between the reconstructed CO₂ concentrations and the national background value represent the combined deviations of ecological fluxes, transport fluxes, and anthropogenic emissions within each grid cell relative to the national average, thereby reflecting the spatial characteristics of carbon sources and sinks [27]. Based on the seamless TAF-XCO₂ dataset, gridded XCO₂ anomalies (ΔXCO₂) over China were derived by subtracting the annual mean national background value from each grid cell, enabling an analysis of the overall distribution patterns and temporal variations of carbon sources and sinks across the country. Concurrently, the CAMS-XCO₂ reanalysis product for the same period was incorporated as a comparative validation dataset to further assess the consistency of the results.

Based on two satellite retrieval datasets, TAF-XCO₂ and CAMS-XCO₂, this study calculated gridded XCO₂ anomalies (ΔXCO₂) across China from 2015 to 2020, aiming to reveal the spatiotemporal variation characteristics of carbon sources and sinks in the China region. Figure 2 shows the multi-year average spatial distribution pattern of TAF ΔXCO₂, and Figure A1 presents that of CAMS ΔXCO₂. The two data products exhibit good consistency: significant positive anomalies are observed in the Beijing–Tianjin–Hebei region, the Yangtze River Delta, and Central China, whereas negative anomalies are found in Northeast and Southwest China. However, numerically, TAF-XCO₂ is approximately 0.53 ppm lower than CAMS-XCO₂ in the Beijing–Tianjin–Hebei region but significantly higher in most parts of the western region. Spatially, significant discrepancies are predominantly concentrated in western regions such as Xinjiang and Tibet. These differences likely stem from fundamental distinctions in the underlying data sources, data assimilation schemes, and reconstruction frameworks employed by the two products. Furthermore, the coarser spatial resolution of CAMS XCO₂ may effectively smooth out localized variability, whereas the TAF-XCO₂ dataset exhibits markedly higher sensitivity to fine-scale regional heterogeneity.

Figure 2. Spatial distribution of average atmospheric ΔXCO₂ in China from 2015 to 2020.

The seasonal ΔXCO₂ values for winter and summer were calculated by averaging the ΔXCO₂ data over the periods of December–January–February (DJF) and June–July–August (JJA), respectively [28]. On a seasonal scale, Figure 3 illustrates the seasonal spatial patterns of ΔXCO₂ retrieved from TAF XCO₂ between 2015 and 2020, and Figure A2 shows those from CAMS XCO₂. In winter, due to weakened vegetation photosynthesis in the Beijing–Tianjin–Hebei region and Central China, the ecosystem’s CO₂ absorption capacity reaches its minimum. The spatial distribution of ΔXCO₂ is primarily dominated by fossil fuel emissions from heating activities [29]. Shandong Province, western Hebei Province, and Henan Province exhibit pronounced positive anomaly areas, with ΔXCO₂ exceeding 2 ppm. Between 32°N and 50°N, ΔXCO₂ is generally positive, while in summer it turns negative, with some areas below −1 ppm, further confirming the seasonal contribution of winter heating emissions. In summer, the strongest negative ΔXCO₂ anomalies appear in the relatively high-latitude areas of China. These regions, characterized by lower intensity of human activities coupled with vigorous vegetation regrowth leading to stronger CO₂ uptake, result in significantly negative ΔXCO₂ [30]. Overall, ΔXCO₂ in the China region exhibits a general pattern of “higher values in winter and lower values in summer” [31], with the exception of the Southwest region, which shows negative anomalies in winter and positive anomalies in summer. The seasonal distribution patterns described above are largely consistent with those derived from CAMS-XCO₂. It is worth noting that CAMS-XCO₂ ΔXCO₂ values in summer are generally higher in areas east of the Heihe–Tengchong Line, and its decreasing trend with latitude is smoother, which may be related to smoothing effects resulting from its coarser spatial resolution [32].

Figure 3. Spatial distribution of seasonal ΔXCO₂ from 2015 to 2022: (a) Winter (b) Summer.

To quantitatively assess the spatial coupling characteristics and interannual evolution between regions of high anthropogenic carbon emission activity and areas of elevated atmospheric CO₂ concentration over China, this study utilized the EDGAR carbon emission dataset and the TAF-XCO₂ product. Grid cells ranking in the top 20% for both emission intensity and XCO₂ concentration were designated as emission hot pixels and concentration hot pixels, respectively. This classification was implemented within ArcGIS 10.8 by applying quantile-based symbology to extract the uppermost 20% of grid cells, followed by a spatial overlay analysis. To facilitate this analysis, the TAF-XCO₂ data were resampled to match the spatial resolution of the EDGAR dataset. As presented in Table 2, the spatial overlap between the two hot pixels categories ranges from 55.49% to 69.09%, indicating that areas of high CO₂ concentration statistically encompass approximately two-thirds of the major ground-level emission sources in most years, thereby demonstrating a reasonable capacity for macroscopic source identification.

Table 2. Spatial overlap between EDGAR carbon emission hot pixels and CO₂ concentration hot pixels in China from 2015 to 2022.

Year	2015	2016	2017	2018	2019	2020	2021	2022
Overlap ratio	62.87%	63.70%	69.09%	66.23%	55.49%	66.76%	66.57%	66.99%

At the regional scale, the “emission–concentration” relationship exhibits significant spatial heterogeneity, leading to considerable differences in the applicability of methods for identifying emission sources based on CO₂ concentration hotspots across different regions [33,34]. As shown in Figure 4, Central China, East China, and North China demonstrate a high degree of “emission–concentration” coupling, with overlap ratios exceeding 80%. Here, concentration signals can effectively reflect the pattern of ground-level emissions. In contrast, the overlap ratios in Southwest China, Northeast China, and South China are less than 30%, where CO₂ concentration hotspots struggle to reliably capture emission sources. This may be related to factors such as more dispersed emission source distributions, favorable diffusion conditions, or stronger regional carbon sink effects [35]. Northwest China represents a moderate coupling area, with an overlap ratio of approximately 50%, but exhibits substantial interannual variability, reflecting the complexity of the emission–concentration relationship in this region.

Figure 4. Proportion of carbon emission overlap by region in China from 2015 to 2022.

It is noteworthy that in 2019, the overlap ratio decreased across almost all regions to varying degrees, with the most pronounced declines observed in Southwest China, South China, and Northeast China. However, Central China and East China still maintained high overlap levels above 83%, demonstrating the greater resilience of their emission–concentration coupling relationships. These regional differences suggest that the strength of the “emission–concentration” relationship is jointly constrained by spatial patterns and atmospheric diffusion conditions [36]. In areas with highly agglomerated emissions, tracing carbon emission sources using CO₂ concentration data possesses high credibility. Conversely, in low-coupling regions, concentration signals must be interpreted cautiously, incorporating regional characteristics.

3.2. EOF Analysis of National Grid Anomalies

While spatial overlay analysis can elucidate the spatial coupling between anthropogenic emission hotspots and elevated XCO₂ to a certain extent, it is inherently limited in capturing the dynamic interactions and temporal variations among distinct driving factors. To address this limitation, this section further introduces Empirical Orthogonal Function (EOF) analysis. This approach decomposes the ΔXCO₂ spatiotemporal matrix into a set of mutually orthogonal canonical spatial modes and their corresponding temporal coefficients, thereby enabling an effective separation of the independent contributions and temporal evolution of multiple factors—including anthropogenic emissions, monsoon-driven transport, and vegetation phenology—throughout the long-term time series.

EOF decomposition was applied to the ΔXCO₂ data over China from 2015 to 2022, revealing four dominant modes. As shown in Figure 5, the first mode exhibits a spatial pattern characterized by “positive anomalies in the southeast and negative anomalies in the northwest,” closely aligning with population density and economic activity intensity. This intuitively reflects the spatial trade-off relationship between anthropogenic carbon emissions and natural ecosystem uptake. Its corresponding time coefficient (PC1) was predominantly in a negative phase before 2019, indicating relatively dominant natural uptake or limited growth in anthropogenic emissions. However, a significant turning point occurred in 2020, with PC1 entering a sustained positive phase. This critical shift may be closely related to a phase transition of the Pacific Decadal Oscillation (PDO) [37] and the decline in aerosol optical depth resulting from air pollution control measures [38]. Together, these factors have made the anthropogenic emission signal increasingly prominent in regional XCO₂ anomalies.

Figure 5. Results of empirical orthogonal function (EOF) analysis for ΔXCO₂ over China from 2015 to 2022.

The second mode manifests as a dipole structure with opposing phases between the north and south. Its formation is closely linked to the interannual variability of the East Asian summer monsoon [39]. PC2 exhibits strong interannual oscillations; for instance, a strong negative anomaly occurred in 2018 (following a weak El Niño), corresponding to a more southerly summer monsoon and enhanced net carbon sinks in the south. Conversely, a strong positive anomaly appeared in 2021 (a strong La Niña year), associated with intensified monsoon circulation and altered hydrothermal conditions in the north [40]. Essentially, this mode results from the monsoon’s influence on surface carbon flux dynamics by modulating regional hydrothermal patterns [41].

The third and fourth modes contribute relatively less to the total variance but reveal important processes at local scales. The third mode concentrates on anomalous signals in the Beijing–Tianjin–Hebei region and Yunnan Province. Its isolated high value in 2017 may be related to stagnant weather events and the transport of cross-border biomass burning emissions [42]. The fourth mode presents a spatial structure characterized by “negative in the southeast and positive in the northwest,” demarcated by the “Hu Line.” This can be viewed as a fundamental geographical background field reflecting differences in underlying surface properties and human activities between eastern and western China.

Notably, as illustrated in Figure 6, the period 2020–2021 constitutes a critical window characterized by dual-mode synergistic enhancement. During this period, the trend reversal in PC1 coincided with a strong positive phase in PC2. This simultaneous occurrence led to the superposition of positive anomalies from the first and second modes over North China, eliciting strong positive XCO₂ anomalies [43]. In contrast, in South China, the first mode was positive while the second mode was negative, resulting in partial cancelation and less significant anomalous signals. This framework provides a mechanistic explanation for the observed high XCO₂ anomalies in eastern China during this period. The turning point in PC1 may be coupled with the complex influences of post-pandemic economic recovery and energy structure transformation [44,45], while the oscillation in PC2 is likely primarily influenced by the El Niño–Southern Oscillation (ENSO).

Figure 6. Spatial phase diagram of PC1–PC2 modes.

3.3. Regional Anthropogenic Emission Increments of XCO₂

3.3.1. Anthropogenic XCO₂ Emission Increments in Urban Agglomerations

We focused on regional XCO₂ anomalies caused by anthropogenic emissions within China’s urban agglomeration areas, selecting the three major urban agglomerations of Beijing–Tianjin–Hebei (BTH), the Yangtze River Delta (YRD), and the Pearl River Delta (PRD) as typical anthropogenic emission source areas. These three regions serve as the core engines of China’s economic development, accounting for over 40% of the national economic output [46], and their carbon emission characteristics are representative for understanding China’s regional carbon cycle and anthropogenic impacts.

Based on the LISA framework introduced in Section 2.2.3, we first calculated the Local Moran’s I for TAF-XCO₂ data across the three major urban agglomerations. A significance level of p ≤ 0.05 was adopted, corresponding to a 95% confidence interval (critical value |Z| = 1.96). Only spatial units whose Local Moran’s I Z-score exceeded this statistical threshold were assigned to one of the four spatial association categories, thereby ensuring the robust identification of high–high (HH) clusters indicative of regional XCO₂ anomalies [47]. Concurrently, a dynamic background concentration curve was constructed by applying a quadratic polynomial fit to the latitudinal means of low–low (LL) cluster values. Local anthropogenic CO₂ increments (dXCO₂) were subsequently derived using Equations (1)–(8).

As summarized in Table 3, the multi-year mean dXCO₂ values attributable to anthropogenic emissions for the period 2015–2019 were 3.46 ± 0.45 ppm for the Beijing–Tianjin–Hebei (BTH) region, 1.30 ± 0.36 ppm for the Yangtze River Delta (YRD), and 0.08 ± 0.14 ppm for the Pearl River Delta (PRD) (uncertainties expressed as 1σ). These results reveal a distinct spatial pattern wherein the anthropogenic emission contribution from the northern heavy industrial region is markedly higher than that of the southern coastal economic zones [48,49], with notably greater uncertainty observed in the PRD—a region characterized by a relatively low emission–concentration overlap ratio. In terms of seasonal variation, winter mean dXCO₂ values for the three regions were 0.60 ppm, 0.78 ppm, and 0.61 ppm, respectively, while summer mean values were 0.94 ppm, 0.16 ppm, and 0.50 ppm. Further analysis indicates a pronounced enhancement of the anthropogenic contribution during summer in the BTH region. This pattern not only reflects persistently high emission intensity during the warm, high-energy-consumption season but may also be associated with the region’s distinctive vegetation cover characteristics—specifically, the vigorous summer vegetation regrowth may lead to an underestimation of background CO₂ concentrations, thereby amplifying the apparent increment attributed to anthropogenic emissions [50].

Table 3. Regional characteristic statistics for the Beijing–Tianjin–Hebei, Yangtze River Delta, and Pearl River Delta regions before and during the COVID-19 pandemic (Unit: ppm).

	2015–2019			2020–2022
	BTH	YRD	PRD	BTH	YRD	PRD
Annual mean CO2	406.26 ± 0.51	406.402 ± 0.36	405.9 ± 0.14	416.25 ± 0.63	416.6 ± 0.47	416.28 ± 0.28
Annual mean dXCO2	3.46 ± 0.45	1.30 ± 0.36	0.08 ± 0.14	2.28 ± 0.73	1.16 ± 0.43	0.28 ± 0.27
Winter mean CO2	407.78 ± 0.85	407.65 ± 0.88	406.53 ± 0.3	418 ± 1.04	418.02 ± 0.76	417.28 ± 0.43
Winter mean dXCO2	0.6 ± 0.6	0.78 ± 0.63	0.61 ± 0.27	2.92 ± 0.74	1.81 ± 0.66	0.4 ± 0.4
Summer mean CO2	403.53 ± 0.58	404.79 ± 0.51	405.48 ± 0.37	413.42 ± 0.74	415.2 ± 0.46	415.65 ± 0.31
Summer mean dXCO2	0.94 ± 0.51	0.16 ± 0.45	0.5 ± 0.39	1.33 ± 0.56	0.49 ± 0.43	0.61 ± 0.3

Dividing the study period into pre-pandemic (2015–2019) and pandemic (2020–2022) phases, comparative statistics for regional mean CO₂ concentration and mean dXCO₂ were compiled and are summarized in Table 3. The data show that during the pandemic, the annual mean dXCO₂ in the BTH region decreased from 3.46 ppm to 2.28 ppm, and in the YRD from 1.30 ppm to 1.16 ppm. This phenomenon indicates a reduction in human activity intensity in these regions during the pandemic, consistent with global trends [51,52]. Conversely, the PRD experienced an increase from 0.08 ppm to 0.28 ppm, reflecting changes in regional emission structures and a trend of enhanced local anthropogenic emissions.

Comparison of winter dXCO₂ further reinforces this spatial differentiation: during the pandemic, the mean winter dXCO₂ in the BTH region reached 2.92 ppm, significantly higher than in the YRD (1.81 ppm) and PRD (0.40 ppm). This highlights that the northern industrial zone maintained relatively high emission levels even under the impact of the pandemic, demonstrating greater continuity and emission inertia in anthropogenic activities [53].

Supported by spatial autocorrelation analysis, High–High (HH) agglomeration areas of XCO₂ anomalies were identified in each region. As shown in Figure 7, during the period 2015–2019, anomalous high-value areas in the BTH region were mainly distributed in Handan and Xingtai cities, located at the intersection of Hebei, Shanxi, Henan, and Shandong provinces, with local increments ranging from 2.95 to 5.74 ppm. The industrial structure of this area is dominated by heavy industries characterized by high pollution and high energy consumption, where vegetation uptake plays a negligible regulatory role, and anthropogenic emissions are predominant [54]. In the YRD, anomalous high-value areas were concentrated in Suzhou, Wuxi, Shanghai, and Jiaxing, with increments between 1.26 and 2.77 ppm, indicating the significant contribution of high-density economic activities to carbon emissions. Anomalous high-value areas in the PRD were primarily located in Shenzhen, Zhongshan, Zhuhai, Dongguan, and southern Jiangmen, with increments ranging from 0.06 to 1.05 ppm, corresponding to its high-intensity urban activities and energy consumption.

Figure 7. Monitoring results of local XCO₂ increments in HH (High–High) clusters of different urban agglomerations from 2015 to 2019.

As shown in Figure 8, entering the 2020–2022 period, the high-value areas in the BTH region remained concentrated in Handan and Xingtai, with increments slightly decreasing but still maintained between 1.79 and 5.57 ppm, suggesting that despite the pandemic’s impact, the fundamental emission pattern did not undergo radical change. In the PRD, the center of gravity for high-value areas shifted towards Guangzhou and Foshan, with increments rising to a range of 0.26–1.12 ppm. This may be associated with the surge in production of pandemic prevention materials and the consequent rise in energy consumption in related industries during the health crisis [55]. The distribution of high-value areas in the YRD showed relatively minor changes, with increments ranging from 0.81 to 2.44 ppm, indicating that this region was less severely affected by the pandemic and its economic activities possessed a certain degree of resilience.

Figure 8. Monitoring results of local XCO₂ increments in HH (High–High) clusters of different urban agglomerations from 2020 to 2022.

In summary, the XCO₂ anomalies in the BTH, YRD, and PRD urban agglomerations exhibit significant regional differences and temporal evolution characteristics, revealing the complex modulating effects of varying economic development models, industrial structures, and pandemic disruptions on anthropogenic carbon emissions [56]. These findings provide important empirical evidence for understanding the spatial patterns and driving mechanisms of carbon emissions in China’s urban agglomeration regions.

3.3.2. Response of Anthropogenic Emission Increments During the COVID-19 Lockdown Period

Since December 2019, the novel coronavirus (COVID-19) pandemic has spread across numerous countries globally. To contain the virus, China implemented stringent lockdown measures that significantly reduced the intensity of human activities, consequently leading to a substantial decline in CO₂ emissions [57,58]. This reduction was particularly pronounced in the early stages of the lockdown. Research by Wang Rong et al. indicated that by the end of February 2020, China’s CO₂ emissions had decreased by approximately 23% year-on-year [57]. Buchwitz et al. conducted a preliminary investigation into XCO₂ concentration changes in East China, observing a decrease of roughly 0.1–0.2 ppm [59]. Nevertheless, systematic quantitative analyses of changes in anthropogenic CO₂ emission increments throughout the entire lockdown period remain insufficient, underscoring a critical research gap.

To address this, our study compared regional XCO₂ anomalies between January and April 2020 and the corresponding months in 2015–2019. The results indicate that, compared to the same period in previous years, XCO₂ anomalies in the YRD region showed a marginal decrease of approximately 0.07 ppm. In contrast, no significant decline was observed in the BTH and PRD regions. The resilience in the BTH region is likely attributable to its rigid industrial structure, dominated by thermal power, steel, and combined heat and power generation, which was less susceptible to pandemic-induced disruptions. Furthermore, extended periods of residential stay during the lockdown likely increased energy consumption for winter heating [60], thereby offsetting emission reductions from the transportation sector. In the PRD region, the sustained operation of core urban functions in these high-density cities generated rigid carbon emissions, particularly evident in the strengthening of the “water–energy–carbon” nexus [61].

To further elucidate the differential responses of dXCO₂ to emission reduction measures across the Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD), and Pearl River Delta (PRD) regions, this study compared dXCO₂ anomalies during 2020 with those of the corresponding months in the 2015–2019 baseline period. As shown in Figure 9 and Figure 10, the results reveal a nuanced pattern: in the YRD region, the overall dXCO₂ difference was approximately −0.01 ppm, yet a positive anomaly of 0.1 ppm persisted in areas of intensive human activity (High–High clusters), indicating that emissions in these core zones remained unabated even as regional totals declined. In the BTH region, overall dXCO₂ increased by 0.13 ppm, with High–High clusters exhibiting a substantial rise of 0.37 ppm, corresponding to increases of 22% region-wide and 45% within the clusters. The PRD region also experienced increases, with overall dXCO₂ rising by 0.07 ppm and High–High clusters by 0.09 ppm, reflecting gains of 23% and 15%, respectively. These divergent patterns underscore the considerable regulatory influence of spatial heterogeneity in industrial structure, energy consumption patterns, and human activity intensity on CO₂ emission dynamics [61].

Figure 9. Monitoring results of local XCO₂ increments in HH (High–High) clusters of different urban agglomerations during the lockdown period from January to April 2020.

Figure 10. Monitoring results of local XCO₂ increments in HH (High–High) clusters of different urban agglomerations from January to April, averaged over 2015–2019.

Furthermore, this study evaluated dXCO₂ concentration changes at the provincial administrative level. Within the BTH region, dXCO₂ in Hebei Province increased by 0.12 ppm (approximately 22%), in Tianjin Municipality by 0.11 ppm (approximately 15%), and in background areas by 0.27 ppm (approximately 38%). In the YRD region, dXCO₂ decreased by 0.02 ppm (approximately 3%) in Jiangsu Province, by 0.07 ppm (approximately 18%) in Anhui Province, by 0.1 ppm (approximately 41%) in Zhejiang Province, and by 0.29 ppm (approximately 30.57%) in Shanghai Municipality. The inter-provincial variations in dXCO₂ change magnitude are likely closely associated with differences in industrial composition, population mobility patterns, and energy consumption characteristics [62].

4. Discussion

Accurately detecting anthropogenic CO₂ increments in urban agglomerations is a scientific foundation for achieving regionally differentiated emission reduction. However, this study indicates that this detection process still faces multiple challenges related to mechanisms of natural interference and regional heterogeneity, which affect the reliable identification of incremental signals.

First, the inherent uncertainty of the data source constrains, to a certain extent, the applicability of the method in specific regions. During model training, the OCO-2 and OCO-3 satellite observations corresponding to TCCON sites were not excluded a priori; instead, a random partitioning strategy was adopted for constructing the training and validation sets in the three-fold cross-validation procedure, which introduces a degree of opacity. Moreover, not all satellite data overlying the TCCON sites were removed. This data handling strategy may give rise to certain errors. Although the TAF-XCO₂ dataset employed in this study exhibits relatively high accuracy over eastern urban agglomerations, the discrimination of grid-scale XCO₂ anomalies in western China remains subject to non-negligible uncertainty due to the absence of effective constraints from TCCON ground-based validation stations. Spatial overlay analysis further reveals pronounced interannual fluctuations in the degree of spatial correspondence between EDGAR emission inventory hot pixels and CO₂ concentration hot pixels at the national scale. This phenomenon may be attributed, on the one hand, to the inherent limitations of relying on a single emission inventory dataset and, on the other, to the significant modulation exerted by large-scale environmental factors such as variations in atmospheric diffusion conditions and interannual variability of natural carbon sinks [63]. The above findings indicate that hot pixel detection methods predicated solely on concentration fields possess certain limitations in regions characterized by spatially dispersed emission sources or favorable atmospheric diffusion conditions.

Second, anthropogenic increments exhibit significant seasonal perturbation characteristics. The dXCO₂ values in all three major urban agglomerations show pronounced seasonal differences, which are not entirely consistent with seasonal expectations of anthropogenic emissions. Taking the Beijing–Tianjin–Hebei region as an example, the summer increment is higher than the winter increment, presenting a phenomenon of “higher values in summer and lower values in winter.” This may be related to improved boundary layer diffusion conditions in summer and the enhanced uptake by vegetation carbon sinks in background areas, which lowers background concentrations. This reflects the difficulty in completely eliminating the seasonal dynamic influence of natural carbon sinks from current LL background areas, and separating vegetation signals from observational data remains a technical challenge [64].

Furthermore, regional industrial structure exerts a profound influence on the response mechanisms of anthropogenic emissions [65,66,67]. The differential dXCO₂ responses during the COVID-19 lockdown period further validate this judgment. In particular, the increase in HH areas of the Beijing–Tianjin–Hebei region reached as high as 45%, reflecting the superimposed effect of rigid demand from heavy chemical industries and heating requirements; the increase in the Yangtze River Delta was only 16%, reflecting the greater emission resilience of its core cities; while the Pearl River Delta experienced an overall increase, possibly related to the surge in production of pandemic prevention materials. This finding indicates that under short-term anthropogenic perturbations, the CO₂ increment response of urban agglomerations is highly constrained by local industrial structure characteristics. Future research should introduce region-specific factors to improve the applicability of unified models.

Regarding the independence of dXCO₂ and methodological limitations, it is important to note that the TAF-XCO₂ seamless dataset used in this study to extract anthropogenic increments incorporated fossil fuel emissions and NO₂ as environmental covariates during the reconstruction process. Consequently, the resulting spatial distribution field of XCO₂ unavoidably embeds some spatial patterns derived from the prior emission information. As demonstrated by Lei, Kiel, and their colleagues, spatial biases in prior emission estimates have a non-negligible influence on both simulated and observe d XCO₂ enhancements [68,69,70]. Therefore, the dXCO₂ derived using the LISA method in this study should be interpreted as a composite indicator that integrates satellite observations with the spatial patterns of prior emissions, rather than a pure observation-based increment entirely independent of emission inventories. Although this indicator has advantages in capturing the spatial structure and relative intensity of anthropogenic emissions, its absolute magnitude is, to some extent, influenced by the spatial allocation of the input prior information. Nonetheless, the annual modeling strategy adopted in this study means that the interannual variability of XCO₂ depends more strongly on changes in the OCO satellite measurements themselves, which alleviates this limitation to a certain degree. In future work, we will further introduce purely observation-based fusion products that do not rely on emission priors, or statistical interpolation products based on in situ measurements, as benchmarks to more thoroughly and scientifically isolate the spatial anthropogenic increment signal.

Finally, it should be noted that this study did not explicitly quantify the spatiotemporal propagation of reconstruction errors when applying the TAF-XCO₂ dataset to EOF and LISA analyses. Validation against TCCON ground-based stations only reflects accuracy at the point scale, whereas the uncertainty in dXCO₂, as a differential quantity, is susceptible to both the selection strategy of the background field and the accumulation of errors. Consequently, the quantitative values of dXCO₂ should be interpreted with caution in western regions where training samples are sparse. Future investigations will incorporate ensemble simulations or Bayesian frameworks to systematically assess the potential interference of these uncertainties on downstream analytical conclusions.

Accurately identifying anthropogenic CO₂ increments in urban agglomerations remains challenged by seasonal interference and regional heterogeneity. Future efforts should focus on multi-source data fusion and the integration of Gaussian plume models with WRF-Chem simulations to establish localized background field models. Such advances will enhance the precision of increment detection and improve the generalizability of the analytical framework across diverse regions.

5. Conclusions

This study systematically analyzed the spatiotemporal variation characteristics of XCO₂ over China based on the TAF-XCO₂ dataset from January 2015 to December 2022, and further identified regional XCO₂ anomalies induced by anthropogenic emissions.

The results indicate that at the national scale, XCO₂ anomalies exhibit a spatial pattern characterized by “positive anomalies in the southeast and negative anomalies in the northwest,” which is highly consistent with population density distribution and economic activity intensity. On a seasonal scale, XCO₂ displays a periodic pattern of “higher values in winter and lower values in summer,” reflecting the combined effects of enhanced emissions from winter heating and increased vegetation photosynthetic uptake in summer. EOF decomposition further extracted four dominant modes: the first mode represents the spatial trade-off relationship between anthropogenic emissions and natural uptake, with its time coefficient undergoing a significant turning point around 2020, possibly related to a phase transition of the Pacific Decadal Oscillation and the decline in aerosol optical depth; the second mode exhibits a north–south dipole spatial structure, significantly modulated by the interannual variability of the East Asian summer monsoon; the third and fourth modes correspond to local-scale influences from localized emission events and the fundamental geographical background, respectively.

In the regional-scale analysis, we proposed a method for identifying regional anthropogenic CO₂ increments based on Local Indicators of Spatial Association (LISA), successfully extracting anthropogenic emission hotspot areas and their background concentrations for the three major urban agglomerations of Beijing–Tianjin–Hebei, the Yangtze River Delta, and the Pearl River Delta. The results show that during 2015–2019, the multi-year mean anthropogenic CO₂ increments (dXCO₂) for the three urban agglomerations were 3.46 ± 0.45 ppm, 1.3 ± 0.36 ppm, and 0.08 ± 0.14 ppm, respectively, revealing a spatial distribution pattern where the northern heavy industrial region exhibits significantly higher values than the southern coastal economic zones. During the 2020–2022 period, dXCO₂ in the Beijing–Tianjin–Hebei region decreased to 2.28 ± 0.73 ppm, slightly decreased to 1.16 ± 0.43 ppm in the Yangtze River Delta, and increased to 0.28 ± 0.27 ppm in the Pearl River Delta, reflecting the differential responses of regional industrial structures and emission resilience to pandemic disruptions. It is noteworthy that in some years, summer anthropogenic CO₂ increments were even higher than those in winter, indicating that current estimation methods based on background fields still struggle to completely eliminate the seasonal dynamic influence of natural carbon sinks. During the COVID-19 lockdown period, compared to historical periods, anthropogenic CO₂ increments showed a slight decrease in the Yangtze River Delta, while slight increases were observed in the Beijing–Tianjin–Hebei and Pearl River Delta regions, suggesting that differences in regional industrial structure are key factors influencing anthropogenic CO₂ increments.

This study provides a new analytical approach for identifying anthropogenic emission responses at national and regional scales using spatiotemporally continuous XCO₂ datasets. The findings confirm that spatially continuous XCO₂ datasets possess good potential for monitoring long-term XCO₂ changes and identifying regional enhancement signals triggered by anthropogenic emissions, offering data support and methodological references for future greenhouse gas emission reduction assessments and policy formulation.