Landscape Patterns and Carbon Emissions in the Yangtze River Basin: Insights from Ensemble Models and Nighttime Light Data

Abstract

Land use patterns are a critical driver of changes in carbon emissions, making it essential to elucidate the relationship between regional carbon emissions and land use types. As a nationally designated economic strategic zone, the Yangtze River Basin encompasses megacities, rapidly developing medium-sized cities, and relatively underdeveloped regions. However, the mechanism underlying the interaction between landscape patterns and carbon emissions across such gradients remains inadequately understood. This study utilizes nighttime light, land use and carbon emissions datasets, employing XGBoost, CatBoost, LightGBM and a stacking ensemble model to analyze the impacts and driving factors of land use changes on carbon emissions in the Yangtze River Basin from 2002 to 2022. The results showed: (1) The stacking ensemble learning model demonstrated the best predictive performance, with a coefficient of determination (R²) of 0.80, a residual prediction deviation (RPD) of 2.22, and a root mean square error (RMSE) of 4.46. Compared with the next-best models, these performance metrics represent improvements of 19.40% in R² and 28.32% in RPD, and a 22.16% reduction in RMSE. (2) Based on SHAP feature importance and Pearson correlation analysis, the primary drivers influencing CO₂ net emissions in the Yangtze River Basin are GDP per capita (GDPpc), population density (POD), Tertiary industry share (TI), land use degree comprehensive index (LUI), dynamic degree of water-body land use (K_water), Largest patch index (LPI), and number of patches (NP). These findings indicate that changes in regional landscape patterns exert a significant effect on carbon emissions in strategic economic regions, and that stacked ensemble models can effectively simulate and interpret this relationship with high predictive accuracy, thereby providing decision support for regional low-carbon development planning.

1. Introduction

Land serves as a crucial transporter of carbon emissions [1,2]. According to the Global Carbon Emissions Statistics, the change in land use represents the second largest anthropogenic source of global carbon emissions [3]. Since 1850, cumulative emissions from changes in land use have amounted to approximately 20.5 billion tons of carbon, accounting for almost one-third of total anthropogenic emissions [4]. Human activities such as fossil fuel combustion, industrial production, urban expansion, and economic development all lead to large emissions of carbon dioxide [5,6], which are ultimately reflected in different land use practices. Therefore, exploring the spatiotemporal distribution patterns and driving factors of carbon emissions associated with regional land use change is essential for establishing a scientific basis to promote low-carbon land use strategies and to support sustainable ecological and socio-economic development.

Currently, the factors that influence the impact of land use on carbon emissions are concentrated mainly in energy intensity [7], economic activity [8], population density [9], industrial structure [10], and expansion of construction land [11]. Wang et al. [12] indicate that the key driver of growth in carbon dioxide emissions from energy consumption in China is sustained economic growth. They propose that optimizing the composition of energy consumption and industrial layout is an effective strategy to reduce carbon emissions. Zhang et al. [13] used Changzhou City as a case study to explore the bidirectional interaction mechanism between land use and carbon emissions and found that there is significant spatial heterogeneity in the distribution of land use patterns and transportation carbon emissions between urban and rural areas. The change in land use significantly influences the spatial and temporal patterns and intensity of regional carbon emissions through synergistic effects on multiple spatial scales, revealing the driving forces behind carbon emissions regarding economic growth, energy consumption and land use.

Scientific identification and quantitative analysis of the factors mainly include STIRPAT [14], Logarithmic Mean Divisia Index (LMDI) [15], SBM [16], panel data regression model [17], etc. Meng et al. [18] analyzed the factors that influence carbon emissions from land use in nine provinces of the Yellow River basin based on the LMDI model and showed that factors such as population density and land use structure have a positive effect on carbon emissions from land use. Yang et al. [19] analyzed the results based on the extended STIRPAT model and found that economic, population, energy, and land factors are direct determinants of regional carbon emissions from land use, with land serving as a key driver. Da Costa et al. [20] discussed the characteristics of spatiotemporal variation and the main factors influencing CO₂ in southeastern Brazil based on regression models and pointed out that the average columnar CO₂ concentration is inversely proportional to vegetation and climate parameters, while the change in land use is directly correlated with CO₂ emissions. These methods or models can only reflect the unidirectional causal relationship between independent variables and dependent variables and cannot explain complex nonlinear relationships and interactions between variables [21]. It is also difficult to dynamically capture the spatio-temporal evolution characteristics of carbon emission factors, which in turn affects the effective identification and analysis of long-term trends [22].

In recent years, machine learning has become increasingly popular in carbon emission simulation research. As a data-driven method for autonomous model construction, it exhibits strong robustness to multicollinearity among explanatory variables during model training [23]. Acheampong and Boateng [24] selected nine variables, including economics, energy and R&D investment, to establish carbon emission prediction models for Australia, China, the United States, Brazil, and India based on the ANN model. They noted that R&D investment, population, urbanization rate, and energy are the main factors that affect carbon emissions. Zhang et al. [25] used the ANN-CA and IWOA-LSTM models to simulate changes in land use and variations in surface temperature in Wuhan, exploring the correlation between carbon emissions and surface temperature. Their results showed that in both 2010 and 2020, the R² values of the linear fits for winter and summer exceeded 0.85. Although such algorithms demonstrate excellent adaptation capabilities when processing non-linear data, the optimization process in high-dimensional parameter spaces is relatively complex, which can easily lead to local optima and overfitting [26].

As a representative ensemble learning algorithm, stacking effectively captures complex features of nonlinear time series by coordinating and integrating diverse base learners and leveraging the trade-off between variance and bias [27], providing a new approach to selecting the main predictive variables. Hoxha et al. [28] constructed a stacking ensemble model that incorporates factors such as population, vehicle mileage, and passenger mileage to predict energy consumption related to transportation in Turkey. The results showed an R² as high as 0.99 in the test set, with vehicle mileage and population variables exhibiting high importance scores. Zhang et al. [29] used a hybrid model that combined quadratic decomposition with stacking ensemble learning to predict carbon emissions in Germany, France, and Italy. The findings showed that the prediction accuracy for Italy and Germany improved by 64.56% and 81.66%, respectively, while MAPE and RMSE for France also improved by 56.66% and 59.11%, respectively. They identified the level of industrialization and energy consumption as the main factors. Therefore, the stacking ensemble model can combine the advantages of different base models and compensate for the limitations of a single model in comprehensively capturing potential correlations in data and avoiding overfitting, constructing a meta-model with higher accuracy and stronger robustness, thereby improving the accuracy of carbon emission predictions and enabling effective assessment of variable importance.

This study uses the Yangtze River basin as an example, collecting nighttime light data, land use data, and socioeconomic data from 2002 to 2022, and proposes a CO₂ prediction model based on the stacking ensemble learning algorithm. This model integrates datasets from different spatial scales, using LightGBM, XGBoost, and SVR as base learners. Through comparative analysis, its superiority is demonstrated, and CO₂ prediction maps are generated, revealing the spatio-temporal characteristics of land use and carbon emissions in the Yangtze River basin and assessing the importance of contributing features. The main objective of this paper is to (1) calculate indirect CO₂ emissions using nighttime remote sensing data, analyze regional spatio-temporal evolution characteristics and trends, analyze the driving factors affecting CO₂ net emissions, and explore the relationship between CO₂ net emissions and land use; (2) quantify various indicators of landscape patterns, predict urban CO₂ net emissions using stacking learning algorithms, and analyze feature importance to provide practical references for the scientific formulation of carbon emission reduction measures in the Yangtze River basin.

2. Materials and Methods

2.1. Research Area

The Yangtze River basin (N 24°27′–35°54′, E 90°33′–122°19′) is the largest river basin in China, with a total length of approximately 6300 km. It extends across the eastern, central and western economic zones of the country, encompassing 17 provinces (autonomous regions) and 2 municipalities, with a total land area of approximately 1.8 million km². The region spans a vast area, stretching from the Yellow River and Huai River basins in the north to the Pearl River basin and the Fujian-Zhejiang water system in the southeast. There are significant disparities in economic development and natural differences between the upper, middle, and lower reaches of the region. The terrain slopes from west to east and is characterized by complex and diverse landforms, with a variety of land use types. Therefore, studying its land use status in this basin is of great importance for optimizing land use patterns and promoting sustainable regional development (Figure 1).

2.2. Data Sources and Preprocessing

As shown in

Table 1. Data sources.

Data Name	Time	Data Source	Data Information
Land use data (Version 1.0.3)	2002–2022	The 30 m annual land cover datasets and their dynamics in China from 1985 to 2023 https://zenodo.org/records/12779975 (accessed on 1 October 2025)	30 m × 30 m Cropland, forest, grassland, water, built-up land, unused land (six land use types)
Carbon emissions data (based on fossil fuel data)	2002–2022	ODIAC Fossil fuel emission dataset from the center for Global Environmental Research	1 km × 1 km
Nighttime lighting data	2002–2022	https://eogdata.mines.edu/products/dmsp/ (accessed on 8 November 2024) https://eogdata.mines.edu/products/vnl/ (accessed on 8 November 2024)	DMSP/OLS (2002–2012) NPP/VIIRS (2013–2022)
Socioeconomic data	2002–2022	China Urban Statistical Yearbook, Statistical Yearbooks of Provinces, Municipalities, and Prefectures	Population density, GDP, urbanization rate, and proportion of industry structure

, the data sources for this article mainly include land use data, carbon emissions data (based on fossil fuel data), nighttime lighting data, and socioeconomic data.

The nighttime light data used in this study combines radiance calibrated DMSP/OLS imagery (2002–2013) and NPP/VIIRS imagery (2013–2022). Following the studies of Zhang et al. [30], a series of correction procedures were applied to the DMSP/OLS imagery. Firstly, the raw images underwent standardized projection and resampling, ultimately achieving a uniform spatial resolution of 1 km × 1 km pixels. Subsequently, saturation correction and continuity correction were performed sequentially. For NPP/VIIRS data, the monthly data were first averaged annually to produce year-scale imagery. Subsequently, this imagery underwent projection and resampling, followed by outlier removal and continuity correction, yielding continuous night-time light data for the period 2013–2022. Finally, based on the fitting of 2013 DMSP/OLS and NPP/VIIRS night-time radiance data, the NPP/OLS data were processed to produce continuous night-time radiance data imagery for the period 2002–2022.

2.3. Variable Selection

Considering the characteristics of the study area and the availability of data, this paper selected 17 primary indicators across three aspects (land use structure, landscape index, and socio-economy) as shown in

Table 2. Indicators of factors affecting land use and carbon emissions.

	Indicator Layer		Formula	Instructions
Land use patterns	Land use mix index		$LUM={\displaystyle \frac{-\sum _{i=1}^n{P}_i\ln P_i}{\ln n}}$	$P_i$ is the area proportion of the land use type (%); n is the number of land use types.
	Land use degree comprehensive index		${LUI}_j=100\times {\displaystyle \sum _{i=1}^n}{A}_i\times {\displaystyle \frac{C_i}{C}}$	$A_i$ is the number of land use classification levels, $C_i$, $C$ is the land use area at level i and the total land area in the region.
	Land use dynamics	Single land use dynamic	$K_T={\displaystyle \frac{A_b-A_a}{A_a}}$	$A_a$ and $A_b$ are the areas of land use type at the outset and end of the study period.
		Comprehensive land use Dynamics	${LA}_T={\displaystyle \frac{\sum _{i=1}^n∆L{A}_{ij}}{\sum _{i=1}^n{LA}_i}}\times 100\mathrm{\%}$	$∆{LA}_{ij}$ is the area of land use type $i$ converted to land use type $j$ during the study period; ${LA}_i$ is the area of land use type $i$ at the beginning study.
Landscape index	Number of patches $\left(\mathrm{N}\mathrm{P}\right)$		$NP=N$	NP describes the heterogeneity of the whole landscape.
	Patch density $\left(\mathrm{P}\mathrm{D}\right)$		$PD={\displaystyle \frac{1}{A}}{\displaystyle \sum _{j=1}^M}{N}_i$	The larger the PD, the more dispersed the urban land use.
	Largest patch index ($\mathrm{L}\mathrm{P}\mathrm{I}$)		$LPI={\displaystyle \frac{max\left(a_1,a_2,a_{3,}\dots ,a_n\right)}{A}}\times 100\mathrm{\%}$	LPI is the proportion of the largest patch to the total area of the landscape.
	Interspersion andjuxtaposition index ($\mathrm{I}\mathrm{J}\mathrm{I}$)		$\mathrm{IJI}={\displaystyle \frac{-\sum _{k=1}^m\left[\left(\frac{e_{ik}}{\sum _{k=1}^m{e}_{ik}}\right)\mathrm{l}\mathrm{n}\left(\frac{e_{ik}}{\sum _{k=1}^m{e}_{ik}}\right)\right]}{\mathrm{l}\mathrm{n}\left(m-1\right)}}\times 100$	IJI calculates the overall distribution and juxtaposition of individual patches.
	Landscape shape index ($\mathrm{L}\mathrm{S}\mathrm{I}$)		$LSI={\displaystyle \frac{0.25\sum _{k=1}^m{e}_{ik}^{\mathrm{*}}}{\sqrt{TA}}}$	LSI is used to measure the total length or density of the edge.
	Modified Simpson’s Evenness Index $\left(\mathrm{M}\mathrm{S}\mathrm{I}\mathrm{E}\mathrm{I}\right)$		$MSIEI={\displaystyle \frac{-ln\sum _{i=1}^m{P}_i^2}{\mathit{ln}m}}$	MSIEI is the uniformity of distribution between patch types.
	Contagion index $\left(\mathrm{C}\mathrm{O}\mathrm{N}\mathrm{T}\mathrm{A}\mathrm{G}\right)$		$CONTAG=\left[{\displaystyle \frac{{\sum }_{i=1}^m\sum _{k=1}^m\left[p_i^o\frac{g_{ik}}{\sum _{k=1}^m{g}_{ik}}\right]\left[\ln \left(p_i^o\frac{g_{ik}}{{\sum }_{k=1}^m{g}_{ik}}\right)\right]}{2\ln \left(m\right)}}\right]\times 100$	CONTAG is the connectivity between different types of patches. Higher values indicate higher connectivity.
	Perimeter-Area Fractal Dimension $\left(\mathrm{P}\mathrm{A}\mathrm{F}\mathrm{R}\mathrm{A}\mathrm{C}\right)$		$PAFRAC={\displaystyle \frac{2}{\frac{\left[n_i\sum _{j=1}^n\left({Inp}_{ij}\ast {Ina}_{ij}\right)\right]-\left[\left(\sum _{j=1}^n{Inp}_{ij}\right)\left({\sum }_{j=1}^n{Inaij}_{ij}\right)\right]}{\left(n_i{\sum }_{j=1}^n{Inp}_{ij}^2\right)-{\left({\sum }_{j=1}^n{Inp}_{ij}\right)}^2}}}$	PAFRAC reflects the complexity of the shape on a range of spatial patches (patch sizes).
Social and economic factors	Urbanization Rate ($\mathrm{U}\mathrm{R}$)		Urban population/total population of the region	The proportion of the urban population in the total population of the region.
	Population Density ($\mathrm{P}\mathrm{O}\mathrm{D}$)		Total population/Administrative area	Population density of prefecture-level cities.
	GDP per capita ($\mathrm{G}\mathrm{D}\mathrm{P}\mathrm{p}\mathrm{c}$)		GDP total value/average annual population total	Per capita economic scale and level of development in the region.
	Secondary industry share ($\mathrm{S}\mathrm{I}$)		Secondary industrial output/total industrial output	The proportion of the output value of the secondary industry to the total output value in the region.
	Tertiary industry share ($\mathrm{T}\mathrm{I}$)		Tertiary industry output/total output	The proportion of the output value of the tertiary industry to the total output value of the region.

. Among these, the single land use dynamics indicator is further subdivided into five sub-indicators: cropland (K_crop), woodland (K_wood), grassland (K_grass), water bodies (K_water), and built-up land (K_cons). In total, 21 factors are employed to delineate the influences of land use practices on carbon emissions.

2.4. Methods

2.4.1. Land Use Carbon Emission Estimation Methods

Based on the calculation methods provided in the IPCC, this study divides the carbon emission effects of land use into two types: direct emissions and indirect emissions.

(1)

Direct carbon emissions calculation method

Following existing research, carbon emissions from cropland, forest land, grassland, water areas, and unused land were estimated using the direct emission method. The calculation formula is as follows:

$$E_x=\sum S_i\ast T_i$$

(1)

where $E_x$ is carbon emissions (absorption) from cropland, forest land, grassland, water areas, and unused land. $S_i$ is the area of land type $i$ (ha). $T_i$ is the carbon emission coefficient of land type $i$ (

Table 3. Carbon emission coefficient of land use type (t·hm⁻²·a⁻¹).

Land Type	Carbon Emission Factor	Reference Sources
Cropland	0.4970	Cai et al., 2005 [31]
Forest	−0.6440	Fang et al., 2007 [32]
Grassland	−0.0205	Sun et al., 2015 [33]
Water	−0.0230	Yuan et al., 2019 [34]
Unused Land	−0.0050	Shi et al., 2012 [35]

). $E_x$ is positive for carbon emissions and negative for carbon absorption. The CO₂ emission (absorption) coefficients were established by referencing previous studies and incorporating the actual conditions of the Yangtze River Basin, as shown in Table 3.

(2)

Indirect carbon emissions calculation method

Nighttime lighting data can effectively reflect the intensity of human social activities, and many studies have demonstrated a linear correlation between nighttime light intensity and carbon emissions [36], which can estimate regional carbon emissions with high precision. Based on data released by the ODIAC platform, this study integrates monthly time series observation data to obtain carbon dioxide emissions from fossil fuels data from 2002 to 2022, achieving spatiotemporal inversion of annual indirect carbon emissions during the study period.

$$E_i=\sum M_i$$

(2)

where $E_i$ is indirect carbon emissions. $M_i$ is the monthly carbon emissions that is a grid dataset of CO₂ emissions generated by industrial processes such as fossil fuel combustion and cement production.

(3)

CO₂ net carbon emissions

CO₂ net carbon emissions are the sum of direct and indirect carbon emissions, calculated as follows:

$$E=E_x{+E}_i$$

(3)

where $E_x$ is direct carbon emissions, which are carbon emissions from cropland, forest, grassland, water bodies, and unused land; $E_i$ is indirect carbon emissions, which are carbon emissions generated on construction land.

2.4.2. Land Use and Carbon Emissions Relationship Model

(1)

XGBoost model

XGBoost is a highly efficient and scalable tree boosting system based on a gradient boosting framework proposed by Chen and Guestrin. It constructs models through iterative integration of weak learners, with each iteration employing a gradient descent strategy to fit negative gradient residuals and optimize accuracy [37]. Its core strengths lie in its excellent balance among superior prediction accuracy, robustness, computational efficiency, and model interpretability when processing structured data [38]. XGBoost effectively prevents overfitting by incorporating a regularization term to control model complexity, while simultaneously achieving multi-system optimization to enhance execution speed and efficiency [39].

(2)

CatBoost model

The CatBoost model is a machine learning algorithm proposed by Dorogush et al. [40] that is based on the gradient boosting decision tree (GBDT) framework. This method employs iterative ensemble learning using a symmetric binary tree structure, where each subsequent models optimize the prediction residuals of preceding models to progressively reduce prediction errors. Compared with traditional GBDT approaches, the significant advantage of CatBoost lies in its ability to efficiently and robustly handle categorical features while effectively mitigating gradient bias and prediction bias. Therefore, it reduces the risk of overfitting and enhances the predictive accuracy and generalization performance [41].

(3)

LightGBM model

LightGBM is an efficient machine learning method based on the gradient boosting framework, incorporating histogram decision tree optimization. It is suitable for complex tasks such as regression and classification [42]. Its core mechanisms include gradient unilateral sampling and mutually exclusive feature bundling, both of which significantly reduce data scale and computational complexity while ensuring controllable information loss. Simultaneously, the histogram algorithm accelerates node splitting computations and reduces memory consumption by discretizing continuous features into finite intervals. Additionally, LightGBM employs a leaf-first growth strategy [43], selecting the leaf node with the greatest gain for splitting at each step to achieve higher model accuracy (Figure 2).

(4)

Stacking ensemble learning models

The stacking ensemble learning model enhances overall performance by integrating the predictive capabilities of multiple learners and is fundamentally composed of two layers: base learners and a meta-learner. Its core principle is to generate predictions using multiple primary learners, then feed these results as new features into a secondary learner for training and fusion, thereby producing the final prediction value [44]. In this study, the stacking regression framework was constructed with LightGBM, XGBoost, and Support Vector Regression (SVR) as base learners, and Ridge regression as the meta-learner. This configuration enables complementarity between tree-based models and kernel methods, thereby leveraging the strengths of different learning mechanisms to improve predictive accuracy. Notably, this paper introduces a passthrough strategy within the Stacking framework. This approach preserves the original feature inputs while transmitting the meta-learners’ predictions, enabling the meta-learners to balance both original information and multi-model outputs during weighted fusion to improve the robustness and interpretability of the fusion results.

Furthermore, to ensure optimal performance of each base learner under varying levels of complexity, this paper adopts an automated hyperparameter optimization method for systematic adjustment of core parameters. Through global parameter search and cross-validation mechanisms, an effective balance was achieved between model complexity and generalization capability, significantly enhancing the predictive accuracy and robustness of the final ensemble model [45]. Finally, the SHAP method was applied to interpret the model outputs, providing a quantitative assessment of the contributions of individual features to the predictions. The overall framework of the model is illustrated in Figure 3.

2.4.3. Model Evaluation

The R², RMSE, and RPD are used to evaluate and compare the stability and accuracy of different models, as well as to assess the agreement between predicted and observed CO₂ emissions. A higher R² value, approaching 1, indicates a stronger correlation; the smaller the RMSE, the smaller the error between the predicted value and the actual value, indicating a better model prediction performance. RPD is defined as the ratio of the standard deviation to the RMSE, where a higher RPD value indicates superior predictive capability.

$$\begin{array}{l}{R}^2=1-{\displaystyle \frac{\sum _{i=1}^n{\left(y_i-\widehat{y}\right)}^2}{\sum _{i=1}^n{\left(y_i-\overline{y}\right)}^2}}\end{array}$$

(4)

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

(5)

$$\begin{array}{l}RPD={\displaystyle \frac{SD}{RMSE}}\end{array}$$

(6)

where $y_i$ is the actual CO₂ emission value, ŷ represents the predicted CO₂ emission value, $\overline{y}$ is the average of actual measurements of CO₂, $SD$ is the standard deviation, n is the number of samples.

2.4.4. Model Design

During the process of model construction, this paper employs a method combining time series grouping with time-blocking to ensure temporal independence between model training and validation. Time-blocking is a validation strategy specifically designed for time series data in which the core principle involves partitioning the data based on the temporal sequence of observations. This ensures the training set comprises solely historical data, while the test set consists of observations that are temporally more delayed [46]. This approach effectively mitigates the risk of future information leaking into the model training process, which conventional random partitioning could not guarantee [47]. Accordingly, this paper uses data from 2002, 2007 and 2012 as the training set, comprising 399 samples in total, while data from other years form the test set, comprising 266 samples in total. This represents a training-to-test split ratio of approximately 6:4 (shown in

Table 4. Model data partition.

	Number	Min (Mt)	Max (Mt)	Standard Deviation (Mt)
Training set	399	−2.73	71.94	6.72
Test set	266	−2.35	81.77	9.93

).

3. Results

3.1. Spatio-Temporal Characteristics of the Change in Land Use

There were significant differences in changes between land use types in the Yangtze River basin during the study period (Figure 4). Overall, there has been a significant decline in cropland area, while the areas of construction land and forest land have increased. Specifically, the cropland area decreased from 5.37 × 10⁷ hm² in 2002 to 5.09 × 10⁷ hm² in 2022. The converted cropland area was 1.3 × 10⁶ hm² between 2002 and 2012, and 1.5 × 10⁶ hm² between 2012 and 2022. Among them, forests were the main type of land converted from cropland, covering an area of 3.77 × 10⁶ hm², accounting for 70.17% of the total converted area. Construction land was the second largest type, accounting for 20.7% of the total converted area. The area of construction land has increased from 2.26 × 10⁶ hm² in 2002 to 4.61 × 10⁶ hm² in 2022. From 2002 to 2012, the area of other types of land converted into construction land was 1.19 × 10⁶ hm², and from 2012 to 2022, it was 1.16 × 10⁶ hm². Among these, from 2002 to 2012, the area of cropland converted to construction land was 1.11 × 10⁶ hm²; from 2012 to 2022, the area of such conversion was 1.07 × 10⁶ hm². The forest area increased from 8.28 × 10⁷ hm² in 2002 to 8.50 × 10⁷ hm² in 2022. Overall, the expansion of construction land and forest land was primarily the result of the reduction in cropland (

Table 5. Changes in land use area from 2002 to 2022.

Land Use Type	Area in 2002 (hm2)	Area in 2012 (hm2)	Increase/ Decrease Rate (2002–2012)	Area in 2022 (hm2)	Increase/ Decrease Rate (2012–2022)	Increase/ Decrease Rate (2002–2022)
Cropland	5.37 × 107	5.24 × 107	−2.46%	5.09 × 107	−2.86%	−5.24%
Forest	8.28 × 107	8.37 × 107	1.07%	8.50 × 107	1.55%	2.63%
Grassland	3.41 × 107	3.31 × 107	−2.95%	3.23 × 107	−2.36%	−5.24%
Water	3.83 × 106	3.86 × 106	0.88%	3.46 × 106	−10.39%	−9.60%
Built-up land	2.26 × 106	3.45 × 106	52.20%	4.61 × 106	33.68%	103.46%
Unused land	1.56 × 106	1.78 × 106	14.56%	2.00 × 106	12.40%	28.77%

and Figure 4).

During the study period, changes in land use in the Yangtze River basin exhibited significant regional variations. In the upper Yangtze River basin (UYRB), terrain restrictions limited change in land use, the main transformation being the conversion of grasslands to unused land. However, in regions such as Yunnan and Sichuan, a large amount of cropland has been converted to grassland. As a traditional agricultural core area, the middle Yangtze River basin (MYRB) initially focused on converting forest land and water areas into cropland. However, with the strengthening of ecological protection policies in recent years, which aim to restore the ecological functions of the river basin, a reverse conversion from cropland to forest land and cropland to water areas has been achieved. The Lower Yangtze River basin (LYRB) experienced rapid urbanization and industrialization during the study period, characterized primarily by the conversion of cropland to construction land (Figure 5).

3.2. Spatio-Temporal Characteristics of Carbon Emissions

To estimate net carbon emissions for the Yangtze River basin, this study utilizes corrected DMSP/OLS and NPP/VIIRS night-time total brightness values (TDN) to establish a correlation model between TDN and fossil fuel emissions from the ODIAC dataset. A logarithmic regression was fitted between annual provincial TDN and the corresponding ODIAC emissions for 2002–2022 (

Table 6. Carbon emission prediction models established based on TDN values and indirect carbon emissions in the Yangtze River basin provinces from 2002 to 2022.

Provinces	Regression Results			Provinces	Regression Results
	Fitting Equations	R2	F value		Fitting Equations	R2	F Value
Henan	y = 4636ln(x) − 4399.7	0.93	191.50	Hubei	y = 1763.3ln(x) − 248.69	0.91	160.03
Zhejiang	y = 4531.1ln(x) − 7112.1	0.94	238.67	Chongqing	y = 710.02ln(x) + 668.3	0.92	183.16
Qinghai	y = 336.32ln(x) + 248.42	0.95	337.49	Jiangsu	y = 9486.3ln(x) − 22,398	0.93	230.47
Gansu	y = 2040.8ln(x) − 1357.4	0.92	186.76	Shanghai	y = 5684.6ln(x) − 13,858	0.93	182.26
Xizang	y = 77.545ln(x) + 172.02	0.92	182.58	Jiangxi	y = 1049.9ln(x) + 618.69	0.92	130.34
Guangxi	y = 1378ln(x) + 811.11	0.94	247.83	Guizhou	y = 884.49ln(x) + 1627.9	0.93	177.91
Fujian	y = 2293.2ln(x) − 1667.3	0.95	311.33	Hunan	y = 1400.1ln(x) + 651.89	0.94	237.96
Shaanxi	y = 3591.4ln(x) − 3872.8	0.97	548.33	Yunnan	y = 2049ln(x) − 1059.2	0.92	192.82
Guangdong	y = 10,223ln(x) − 27,931	0.91	149.75	Sichuan	y = 1529ln(x) + 31.222	0.92	171.56
Anhui	y = 2595.4ln(x) − 40.555	0.93	230.72

). The regression results indicate a positive and statistically significant association (p < 0.0001). On the basis of this regression relationship, indirect carbon emissions for each city were estimated by applying the corresponding city-level TDN values. These estimates were then added with the direct carbon emissions to derive the net carbon emissions at the city level, measured in Mt.

CO₂ net emissions in the Yangtze River basin experienced two phases from 2002 to 2022: rapid growth and slow growth, which showed different temporal characteristics (Figure 6). Between 2002 and 2012, carbon net emissions increased rapidly, rising from 207.26 Mt in 2002 to 573.51 Mt in 2012, with an average annual growth rate of 10.71%. Although carbon emissions continued to grow after 2012, the growth rate slowed significantly, reaching 676.01 Mt in 2022, an increase of 17.87% compared to 2012, with an average annual growth rate of 1.66%. In general, CO₂ net emissions in the Yangtze River basin show a clear trend of phased increases, which can be divided into two stages: the first stage is from 2002 to 2012, when carbon emissions increased rapidly; the second stage is from 2012 to 2022, when CO₂ net emissions have fluctuated. Although CO₂ net emissions in individual cities experienced slight increases during the latter period, the general trend indicates a movement toward stabilization.

In terms of spatial distribution, the average CO₂ net emissions in the upper, middle, and lower reaches of the Yangtze River basin in 2022 were 2.54 Mt, 3.32 Mt, and 14.36 Mt, respectively (

Table 7. Interannual variation in mean and standard deviation of carbon emissions in UYRB, MYRB and MYRB (Mt).

		2002	2007	2012	2017	2022
UYRB	Mean	0.70	1.44	2.14	2.18	2.54
	SD	1.88	3.16	4.34	4.43	5.06
MYRB	Mean	1.11	1.97	2.80	2.87	3.32
	SD	1.23	2.03	2.88	2.94	3.37
LYRB	Mean	4.50	8.22	12.46	12.73	14.62
	SD	6.13	11.46	17.21	17.58	19.74

and Figure 7). CO₂ net emissions in Chongqing and Chengdu, located in the UYRB, were 30.02 Mt and 20.10 Mt, respectively, significantly higher than the regional average. The total carbon emissions of the two cities accounted for 32.4% of the total CO₂ net emissions in the UYRB. The CO₂ net emissions of Wuhan, Changsha, and Nanchang in the middle reaches of the Yangtze River basin are also higher than the regional average, which are 18.17 Mt, 11.81 Mt, and 7.51 Mt, respectively. These three cities collectively contributed 24.05% of the total MYRB emissions. In the LYRB, cities such as Shanghai, Suzhou, and Nanjing also recorded CO₂ net emissions that exceeded the regional average, with values of 81.77 Mt, 68.53 Mt, and 27.98 Mt, respectively, accounting for 49.67% of the total emissions in the LYRB.

3.3. Spatial-Temporal Variation and Correlation Analysis of Influencing Factors

The effects of different factors on carbon emissions vary, and their impacts also differ at different stages. This article categorizes the various indicator factors into two groups: natural factors and human factors. Among these, factors with significant human factors include GDPpc, POD, UR, and TI (Figure 8). The minimum GDPpc increased from 101 yuan to 16,793 yuan, representing a growth of more than 66 times, with its spatial distribution expanding from a point to a plane. In terms of spatial distribution, LYRB remains at a high level, while UYRB and MYRB have also grown significantly, with regional differences still existing. The maximum value of POD increased from 2582.49 to 4729.46 people/km², with an increase of 83%. The trend of population concentration in large cities such as Shanghai, Suzhou, and Nanjing has intensified, and regional population density differences remain significant. During the study period, UR increased significantly, with an absolute increase exceeding 20 percentage points. The range of high and medium values expanded, showing dynamic synergistic characteristics of “growth and balance”. TI also increased significantly, with the highest threshold range increasing from 43.90–56.16% to 56.75–74.13%. The process of industrial structuring continues to advance, and spatial heterogeneity is decreasing in the LYRB.

Natural factors that significantly influence carbon emissions include LUI, NP, K_water, LPI, K_grass, and LSI (Figure 9). During the study period, the spatial pattern of LUI has gradually spread from the early highly concentrated areas in the southeast to the northwest, while the intensity differences within the region have tended to narrow. The number of core patches with high NP values (399,274–659,276) showed a relative contraction trend, while patches with medium to low values (118,440–230,204) expanded, forming a continuous spatial development axis. The high-value K_water zone (0.08–0.20) is mainly distributed in a contiguous area in the northwest region, which has decreased in 2022, and the overall trend is converging towards the median and low-value zones. During the study period, the LPI showed an overall upward trend, with the lowest value increasing from 11.56 to 13.23 and the highest value increasing from 96.51 to 97.06, indicating an increase in the dominance of the largest patches in the region. The lower limit of K_grass increased from −0.18 to −0.16, while the upper limit decreased from −0.08 to −0.11, indicating that both the intensity and the extent of the reduction in grassland patches have been weakened. From a spatial distribution perspective, regions with high values have gradually shifted from a scattered distribution to concentration in upstream and midstream areas, showing a clear trend of spatial contraction. During the study period, the high value zones of LSI exhibited a clear trend of shifting from the north toward the southwest, while the overall complexity of ecological land use patterns increased throughout the spatiotemporal evolution of the landscape.

To quantitatively assess the impact of various factors on CO₂ net emissions, Pearson correlation analysis was used to investigate the relationship between CO₂ net emissions and various influencing factors (

Table 8. Correlation coefficient between CO₂ emissions and various indicators.

LUM	LUI	LA	K					GDPpc	POD	UR
			Kcrop	Kwood	Kgrass	Kwater	Kcons
0.264	0.470	0.170	−0.042	0.016	−0.200	−0.101	0.008	0.526	0.836	0.270
SI	TI	NP	LPI	LSI	MSIDI	PD	IJI	CONTAG	PAFRAC
0.105	0.296	0.057	−0.266	0.134	0.308	0.062	0.025	−0.238	−0.048

). The analysis shows that CO₂ net emissions are positively correlated with LUI, LUM, LA, K_wood, K_cons, GDPpc, UR, POD, TI, SI, NP, PD, LSI, IJI, and MSIDI, and negatively correlated with K_crop, K_grass, K_water, LPI, CONTAG, and PAFRAC. Among these, POD and GDPpc are highly correlated with CO₂ net emissions, with correlation coefficients of 0.836 and 0.526, respectively, indicating that densely populated areas and higher economic levels are typically accompanied by higher energy consumption and industrial activity, resulting in increased carbon emissions. Meanwhile, LUI and LUM are also significantly positively correlated with CO₂ net emissions, with correlation coefficients of 0.470 and 0.264, respectively. Additionally, CONTAG and LPI negatively influence CO₂ net emissions, with correlation coefficients of −0.238 and −0.266, respectively. The high values of CONTAG and LPI show that landscape patches are more connected and concentrated, which is conducive to the functioning of land ecosystems. For example, continuous vegetation patches such as forests and water bodies with strong carbon sink capacity can absorb and store carbon dioxide more effectively, reducing the net release of carbon emissions.

3.4. Analysis of the Model Results

The prediction results of the four models, XGBoost, CatBoost, LightGBM, and stacking, are shown in

Table 9. Analysis of the metrics of the machine learning model.

	Training Set			Test Set
Model	R2	RMSE (Mt)	RPD	R2	RMSE (Mt)	RPD
XGBoost	0.71	3.51	1.87	0.66	5.82	1.70
CatBoost	0.72	3.52	1.91	0.67	5.73	1.73
LightGBM	0.69	3.75	1.79	0.57	6.48	1.53
Stacking	0.87	2.43	2.76	0.80	4.46	2.22

. The results indicate that the RPD values of all four models are higher than 1.6, indicating that all four models have a certain ability to estimate CO₂ net emissions. As shown in Figure 10, XGBoost and CatBoost have relatively high R² and RPD values, along with a lower RMSE. However, the LightGBM model performed poorly, with an R² of 0.57, an RMSE of 6.48, and an RPD reduced to 1.53, indicating its predictive accuracy was significantly lower than that of XGBoost and CatBoost. On the contrary, the stacking model demonstrated superior performance, achieving an R² of 0.80, reducing the RMSE to 4.46, and the RPD to 2.22. Among the four models, there are certain differences in the prediction accuracy for high CO₂ net emissions and low CO₂ net emissions. The XGBoost, CatBoost and stacking models perform relatively well with lower dispersion, while the stacking model performs excellently in predicting both low and high emission levels. The results showed that the stacking model significantly outperformed the other three models in the prediction of CO₂ net emissions. The prediction results of the test set are plotted in a measured-predicted scatter regression chart (Figure 10).

4. Discussion

4.1. Model Performance Discussion

This study used XGBoost, LightGBM, CatBoost, and the stacking ensemble model to predict CO₂ net emissions. The results show that the stacking ensemble model outperforms the individual XGBoost, CatBoost, and LightGBM models in estimating CO₂ net emissions, effectively compensating for the limitations of a single algorithm in capturing noise data and spatiotemporal patterns in complex coupled systems. Specifically, the R² and RPD values of the stacking ensemble model were 21.21% and 30.59% higher than those of XGBoost, 19.40% and 28.32% higher than CatBoost, and 40.35% and 45.10% higher than LightGBM, respectively. In terms of RMSE, it was 23.37% lower than XGBoost, 22.16% lower than CatBoost, and 31.17% lower than LightGBM.

The stacking ensemble model deeply integrates the temporal characteristics and multiscale variables of carbon emission data through the multisource feature extraction capabilities of its base learners, significantly enhancing its ability to capture complex data patterns. This is consistent with the research results of Zhang and Cheng [48]. Compared to single models or simple weighted average ensembles, stacking models demonstrate significantly improved accuracy in carbon emission intensity prediction by using meta learners to adaptively calibrate the bias of base models, which can effectively suppress overfitting and reduce the variance of prediction errors. The key evaluation indicators R² and RMSE have both improved significantly, which is consistent with the conclusions of Hu and Li [49]. In terms of computing efficiency, the base learners can be trained independently, making full use of distributed computing resources to reduce overall training time [50]. Moreover, once training is completed, the meta-learner can rapidly generate predictions. Therefore, the stacking ensemble model achieves an organic balance between accuracy and computational efficiency in the field of carbon emission prediction through its diverse combination of base learners and efficient meta-learning mechanisms, demonstrating strong application potential.

4.2. Evaluation of the Results of the Carbon Emission Calculation

Previous studies have widely employed the bookkeeping model and the emission factor method [51]. The bookkeeping model relies on empirical parameters, while the emission factor method is relatively simple to calculate, benefits from easily available data, and is widely applicable in various contexts [52]. Therefore, this paper uses the emission factor method to calculate direct carbon emissions and integrates remote sensing data to calculate indirect carbon emissions, thus improving the accuracy and operability of the calculations to a certain extent.

Accurate assessment of the impact of land use on carbon emissions requires a comprehensive analysis of landscape patterns in the study area. This paper calculates CO₂ net emissions from 2002 to 2022 and finds that, with the increase in carbon emissions, various types of land use have undergone varying degrees of change over the past 20 years. Among these, the area of construction land increased by approximately 1.61 × 10⁶ hm², resulting in 471.52 Mt of carbon emissions, indicating that construction land is the main source of CO₂ net emissions and significantly affects the levels of CO₂ net emissions in the study area, consistent with the results of Zhu et al. [53]. The spatial concentration of construction land has intensified the dominant effect of fossil fuel combustion and industrial processes on CO₂ net emissions. Rapid industrial development has led to significant changes in land use types, which in turn have affected land use structure and population density, ultimately influencing total CO₂ net emissions in the study area. The conclusion is consistent with the research results of Zhang et al. [54]. Meanwhile, studies show that CO₂ net emissions associated with forests, grasslands, unused land, and water bodies are negative (Figure 11), indicating that the expansion of these land types improves the carbon storage capacity of terrestrial ecosystems [55]. Therefore, rational planning and optimization of land use structure, strengthening ecological land protection and restoration, are essential measures to reduce regional carbon emissions and promote low-carbon sustainable development.

4.3. Feature Importance Analysis

4.3.1. Single-Factor Analysis

This study selected 21 potential influencing factors and ranked them based on their contribution to carbon emissions using SHAP feature attribution. The SHAP analysis results for the top 18 factors are presented in Figure 12a. The results indicate that GDPpc, POD, TI, LUI, and K_water rank among the top factors, showing that socioeconomic development levels and land use intensity are the primary drivers of carbon emissions. Secondly, landscape pattern indicators such as LPI, LSI, and NP also demonstrated significant contributions, indicating that landscape patterns and spatial structural changes play a crucial role in regulating carbon emissions. While variables like LA and UR exhibited relatively minor overall contributions, they still showed certain sensitivities under specific conditions.

The SHAP Beeswarm Plot (Figure 12b) further reveals the differences in marginal effects across various feature values. Both high GDPpc and high POD exhibit positive SHAP values, while low values have limited impact. LUI and K_crop exert limited effects on carbon emissions within low-value ranges, exhibiting either mutual or negligible impacts. However, once a critical threshold is exceeded, they transition to sustained positive contributions, revealing the nonlinear driving mechanisms underlying land use transformation [56]. Landscape pattern factors such as LPI, LSI, and NP exhibit complex trends shifting from low positive values toward high values that stabilize or diminish, indicating that landscape fragmentation and aggregation may produce opposing effects at different stages. The SHAP values for K_wood predominantly fall within the negative range, emphasizing that forest dynamics generally function as carbon sinks [57].

4.3.2. Interaction Effect Analysis

In this paper, the SHAP dependency plot is used to reveal the marginal contribution of a single feature to the model prediction at different value levels, and the potential interaction effects with other features. For the selection of interaction features, this paper employs an automated search strategy. These selected features are then presented using color gradients. Compared to the marginal effects of individual features, this method effectively captures the complex interactions between variables within nonlinear models [58]. Therefore, this paper selected socioeconomic factors (GDPpc, TI), land use structure (LUI, LA), and landscape structure (LSI, NP) from the SHAP bar plot for analysis to explore the trends in their SHAP values and the characteristics of interactions with other factors (Figure 13).

Figure 13 reveals the nonlinear characteristics and interaction effects of carbon emission drivers. Figure 13a indicates that as GDPpc increases, the SHAP values rise steadily from low to high levels while maintaining a consistently positive contribution. This reflects that heightened levels of economic development significantly amplify the driving effect on carbon emissions [59], with this positive effect becoming more significant under conditions of higher LSI. Figure 13b indicates that TI exerts a particularly pronounced positive effect on carbon emissions in the low-to-medium value zones, with this influence diminishing progressively in higher-value areas. When landscape concentration is relatively low, the expansion of the tertiary sector exerts a stronger stimulatory effect on carbon emissions. Figure 13d,e reveal that both LUI and LA exhibit critical threshold effects on carbon emissions. LUI significantly increases carbon emissions as land use intensity rises, and this effect is markedly modulated by NP. LA rapidly increases emissions beyond a threshold, exhibiting an interactive reinforcement effect with TI. Figure 13e,f demonstrate that the SHAP values for LSI and NP increase with their own values and exhibit a positive synergistic effect with POD, highlighting the combined impact of spatial patterns and population concentration.

This shows that socioeconomic factors generally exhibit a threshold effect with limited influence at low values, but their impact rapidly intensifies and maintains a positive contribution once the critical point is crossed. Land use structure indicators (e.g., LUI) and landscape pattern factors (e.g., LSI, NP) predominantly exhibit a shift from negative to positive trends, reflecting the regulatory role of optimized landscape structure in carbon emission processes [60]. Therefore, the formation of carbon emissions is the result of the combined effects of economic development, landscape patterns and land dynamics, all of which exhibit nonlinear characteristics.

5. Conclusions

This paper takes the Yangtze River basin as a case study, combining CO₂ net emission data and land use data, and uses a stacking ensemble model to assess the relationship between landscape patterns and CO₂ net emissions. It compares the predictive capabilities of XGBoost, LightGBM, CatBoost, and the stacking ensemble model for CO₂ net emissions, and it evaluates the importance of variables. The main conclusions are as follows:

(1)

CO₂ net emissions showed a significant upward trend from 2002 to 2022, with a net increase of 468.74 Mt. The main reason for this increase was the rapid expansion of construction land, which increased by approximately 1.61 × 10⁶ hm², causing the proportion of carbon emissions to continue to increase. Meanwhile, carbon sources increased, while carbon sinks (grasslands and water bodies) decreased by 5.24% and 9.60%, respectively, accelerating the net increase in carbon emissions. The acceleration of urbanization and industrialization has driven land development and infrastructure construction, accompanied by the degradation of natural ecosystems and a significant weakening of carbon sink functions, thus contributing to the continuous increase in regional carbon emissions.

(2)

Compared to the other three single models, the stacking ensemble model demonstrates superior predictive capability, showing strong feature extraction capabilities and significant advantages in predictive accuracy and stability. Its R², RMSE and RPD values are 0.80, 4.46, and 2.22, respectively, which are better than the XGBoost model by 21.21%, 23.27% and 30.59%, respectively; better than the CatBoost model by 19.40%, 22.16% and 28.32%; and better than the LightGBM model by 40.35%, 31.17% and 45.10%. The stacking model effectively integrates the strengths of different learning models by combining the prediction results of multiple base learners, making full use of multisource feature information, and significantly improving the fitting and generalization capabilities of complex nonlinear relationships.

(3)

The variable feature importance ranking in the stacking ensemble model shows that CO₂ net emissions are influenced by both human factors (GDPpc, POD, and TI) and natural factors (LUI, K_water, LPI, LSI, NP, and K_crop). When combined with Pearson correlation analysis, CO₂ net emissions were positively correlated with LUI, LUM, LA, K_wood, K_cons, GDPpc, UR, POD, TI, SI, NP, PD, LSI, IJI, and MSIDI and negatively correlated with K_crop, K_grass, K_water, LPI, CONTAG, and PAFRAC. Absolute correlation coefficients are generally consistent with the importance evaluation of the characteristics. The most influential factors are GDPpc, POD, TI, LUI, K_water, and LPI, in that order. This demonstrates that socioeconomic activities and landscape patterns influence carbon emissions through multidimensional pathways, and factors including population density, economic level, and land use intensity significantly regulate the spatial distribution and dynamic changes in regional carbon emissions by altering energy consumption structures and ecosystem functions.

From 2002 to 2022, carbon emissions in the Yangtze River basin exhibited significant spatial variations, with a fluctuating upward trend. Among the dominant factors, GDPpc, POD, TI, LUI, K_water and LPI appeared frequently and had a significant impact on the model. Therefore, future urbanization processes should prioritize optimizing landscape patterns, reasonably planning the scale of construction land, and increasing land types with strong carbon sink capacities. It is essential to improve the efficiency of land use in construction while ensuring the protection of forests, grasslands, and water bodies, thus promoting the establishment of a regionally coordinated development model. However, this study also has certain limitations. Although it analyzes the contribution and impact of CO₂ net emissions in multiple dimensions, including land use structure, landscape index, economic level, and population, it does not explore the specific mechanisms underlying each influencing factor. Future research needs to be further deepened.