Spatiotemporal Decoupling of Vegetation Productivity and Sustainable Carbon Sequestration in Karst Ecosystems: A Deep-Learning Synthesis of Climatic and Anthropogenic Drivers

Abstract

Understanding the spatiotemporal dynamics of vegetation net primary productivity (NPP) and its drivers is critical to sustainable land -carbon management, carbon-neutral development, and ecological restoration in fragile karst landscapes. This study proposes a Pearson Correlation—Deep Transformer (PCADT) model that integrates attention mechanisms and geospatial covariates to enhance NPP estimation accuracy in Guangxi, China—a global karst hotspot. Leveraging multisource remote sensing data (2015–2020), PCADT achieves 10.7% higher predictive accuracy (R² = 0.83 vs. conventional models) at 500 m resolution, thereby capturing the fine-scale heterogeneity essential for sustainability planning. The results reveal a significant annual NPP increase (4.14 gC·m⁻²·a⁻¹, p < 0.05), with eastern areas exhibiting higher baseline productivity (1129 gC·m⁻² in Wuzhou) but western regions showing steeper growth rates (5.2% vs. 2.1%). Vegetation carbon sequestration capacity, validated against MOD17A3HGF data (R² = 0.998), demonstrates spatial consistency (east > west), with forest-dominated Wuzhou contributing 6.5 TgC annually. Mechanistic analyses identify precipitation as the dominant climatic driver (partial r = 0.62, p < 0.01), surpassing temperature sensitivity, while bimodal NPP-altitude peaks (300 m and 900 m) and land -use transitions (e.g., forest-to-cropland conversions) explain 18.5% of NPP variability and reduce regional carbon stocks by 4593 tC. The PCADT framework offers a scalable solution for precision carbon management by emphasizing the role of anthropogenic interventions—such as afforestation—in alleviating climatic constraints. It advocates for spatially adaptive strategies to optimize water resource utilization, enhance forest conservation, and promote sustainable land -use transitions. By identifying areas where water -scarcity relief and targeted afforestation would yield the highest carbon returns, the PCADT framework directly supports SDG 13 (Climate Action) and SDG 15 (Life on Land), providing a strategic blueprint for sustainable development in water-limited karst regions globally.

1. Introduction

Net primary productivity (NPP), defined as the amount of organic matter produced by a plant community after autotrophic respiration, is a crucial component of terrestrial ecosystems, reflecting the production capacity of plant communities under natural conditions [1,2]. NPP also plays a key role in determining the carbon sink and source functions of an ecosystem. Therefore, monitoring NPP variations can assist in evaluating environmental changes and carbon budgets in the region, providing a scientific basis for formulating global climate adaptation strategies.

NPP research dates back to the eighteenth century when Ebermayer first estimated forest production [3]. In the 1930s, Boysen-Jensen introduced the concepts of gross and net plant productivity. At that time, NPP measurement methods were predominantly traditional, particularly sampling approaches, which had limited applicability over large areas [4]. In the 1960s, Lieth et al. developed the original primary production model [5]. As technology has advanced, more precise and reliable NPP estimation techniques and high-quality NPP datasets have been developed [6,7]. MOD17A3x is the most commonly used dataset in published NPP studies [8,9,10]. Studies on plant development and environmental monitoring have validated this dataset, with applications to global changes in various regions, including Guangxi, China [11,12]. Consequently, the updated MOD17A3 variant with improved geographical resolution, MOD17A3HGF [13], is used in this study to provide fundamental NPP data.

Recent NPP research has mostly emphasized its temporal characteristics [14,15] and affecting elements [16]. Provinces, cities, districts, and counties are the primary pixel scales at which the temporal trend of NPP is analyzed [17]. The high-altitude distribution of NPP, areas with growing or declining trends, and the degree of increase are the major subjects of the geographical analysis of NPP [18]. However, the majority of investigations have just employed straightforward linear regression techniques [19]. Mann—Kendall examination of trends was employed in this study to examine the NPP temporal fluctuations. This examination is less impacted by outliers and it is not necessary to use sample information to follow a particular pattern to the dispersion. In the hydrology and climate science sectors, it has seen extensive application. The Transformer model was used to forecast the amount of carbon emitted and sequestered. Transformer has produced significant advances in machine translation, particularly with pre-training models based on Transformer architecture, which have been widely utilized in natural language processing with outstanding results. It is even used in computer vision, speech recognition, and other fields.

Natural and human factors influence NPP, which exhibits spatiotemporal dynamics. NPP is specifically influenced by environmental factors, terrain, and human activities, which have been studied using correlational analysis and other approaches [20,21]. Environmental influences can affect NPP both positively and negatively by altering metabolic activity, growth duration, community composition, demographics, and other factors [22]. Through the redistribution of thermal conditions, topography can induce changes in NPP [23]. The impact of human activities on NPP is multifaceted. Ecological regeneration, such as afforestation and converting cropland to forests, can enhance NPP, while overgrazing, urbanization, infrastructure development, and other activities can reduce NPP [24]. Despite numerous interactions between natural and human causes, most research focuses on the impacts of climate change on NPP and is often restricted to a single climatic zone.

The carbon balance is significantly influenced by plants, which are essential components of terrestrial ecosystems. Vegetation is the most significant natural carbon sink, fixing and absorbing carbon through photosynthesis [25]. Enhancing plants’ ability to sequester carbon is a crucial step toward achieving “carbon neutrality” and a key strategy to mitigate the rise in atmospheric CO₂ levels and climate change [26,27]. The four primary techniques currently used to evaluate vegetative carbon sequestration are plot inventories, eddy covariance, atmospheric inversion, and remote sensing technologies. Remote sensing technology is foundational for assessing the processes and pathways of carbon flow between ecosystems and for developing corresponding models to simulate carbon transformation [28]. Remote monitoring has smaller study scales and longer timeframes compared to other approaches, requires less site-specific data, and is more efficient and timely. The anticipated outcomes are more accurate and can better capture the current carbon balance in the study area.

Most researchers use net ecosystem productivity (NEP), which differs from NPP and heterotrophic metabolism, as a fundamental model to calculate carbon sources and sinks in the study area. Previous studies, however, have primarily focused on evaluating the potential of local ecosystems to store and sequester carbon [29]. Moreover, most studies have lacked comprehensive investigations of the entire ecosystem, as they predominantly focused on the carbon sink potential of grasslands and forests [30]. Additionally, there is a lack of information regarding geographical and temporal variations.

Guangxi is an autonomous region encompassing Central Subtropical, Southern Subtropical, and Northern Tropical zones, characterized by diverse landforms. The western region is significantly impacted by soil erosion, while the northern region experiences frequent rainfall and flooding. The southeastern region is a key maize-harvesting area [31,32]. The study area is more representative and typical due to its complex natural conditions [33]. Comparing the spatiotemporal variations in plant NPP, ecosystem carbon storage potential, and their influencing factors is highly valuable. Additionally, Guangxi has undertaken extensive environmental restoration efforts in recent years, and evaluating the temporal and spatial patterns of regional ecological changes in carbon and NPP sink potential is of great significance.

Recent studies have investigated NPP dynamics in ecologically sensitive regions of China. For example, Zhang et al. [25] examined the spatiotemporal variation in vegetation NPP within Chinese ecological function conservation areas, demonstrating that topography and climate jointly control productivity patterns. They found that altitudinal gradients significantly modulate the response of NPP to precipitation and temperature, with lower elevations exhibiting stronger climate sensitivity. Another line of work focuses on NPP in dryland ecosystems, where limited water availability and seasonal climate fluctuations strongly affect productivity. Wang and Liu (2021) analyzed how seasonal temperature and precipitation variations drive dryland vegetation NPP, revealing that phenological shifts—such as earlier green-up and later senescence—substantially buffer NPP against interannual drought [34]. Their results underscore the importance of phenological timing when interpreting NPP trends in semi-arid regions.

National-scale connectivity analysis of forested landscapes has become increasingly prominent, as illustrated by Lin et al. [35]. In their study, the authors applied graph—theoretic methods to map and quantify habitat corridors at the country level, identifying key linkages that enhance biodiversity preservation. Their results underscore the importance of integrated landscape planning to maintain functional connectivity, which directly impacts carbon sequestration potentials by preserving large contiguous forest tracts. Meanwhile, Zhao and Liu [36] performed a fine—scale assessment of ecosystem service flows in a coastal archipelago. They utilized the InVEST suite (Integrated Valuation of Ecosystem Services and Tradeoffs) to map provisioning, regulating, and cultural services across land—sea gradients in the Zhoushan Archipelago. Their findings highlight spatial hotspots where conservation actions yield the greatest multi—service benefits, reinforcing the need to integrate service provision with productivity modeling. By incorporating these recent studies on connectivity and ecosystem services, our research on NPP and carbon sequestration in Guangxi builds upon and complements emerging frameworks for landscape—scale ecological planning. Specifically, insights from forest network connectivity and multi—service valuation inform our discussion of how landscape configuration and service tradeoffs influence NPP patterns and carbon sinks.

Together, these studies highlight the necessity of integrating topographic, climatic, and phenological factors into NPP modeling frameworks. Building on this foundation, our PCADT model combines Pearson correlation analysis and Deep Transformer architecture to capture complex, spatiotemporal dependencies across Guangxi’s diverse landscapes.

Based on the above, we propose three clear hypotheses. (i) Precipitation is the primary climatic driver of inter—annual NPP variability in Guangxi. (ii) Land -use transitions from cropland to forest/grassland significantly increase local NPP. (iii) The PCADT yields smaller MAE and RMSE than ensemble methods (Random Forest, XGBoost) when predicting NPP.

2. Materials and Methods

2.1. Field of Research

Guangxi is located in southern China (Figure 1), with geographic coordinates ranging from 104°28′ to 112°04′ E longitude and 20°54′ to 26°23′ N latitude. The total area is 237,600 square kilometers. The region has a tropical monsoon climate, with an annual average temperature ranging from 17.5 °C to 23.5 °C and annual precipitation ranging from 841.2 mm to 3387.5 mm. The landscape in the northwest is mountainous, while the southeast is relatively flat, featuring diverse landforms such as mountains, plains, hills, and plateaus. Elevation ranges from 0 to 2141 m, with an average elevation of over 800 m. The predominant soil types are red soil, yellow soil, and mountain brown soil, while the primary land use types include forests, grasslands, and cultivated land. As of the end of 2021, Guangxi consists of 14 cities, with a total population of 50.37 million permanent residents. The study area excludes Weizhou Island and Xiyang Island.

2.2. Datasets and Processing

2.2.1. NPP Measured Data

The measured NPP data were obtained from China FLUX (http://www.chinaflux.org/, accessed on 12 August 2023) and the flux data of the Sanjiang Plain Ecological Experiment Station, Chinese Academy of Sciences (http://sjm.cern.ac.cn/, accessed on 12 August 2023).

2.2.2. Remote Sensing Data

The NPP data were obtained from NASA’s EOS/MODIS data (https://lpdaac.usgs.gov, accessed on 12 August 2023), specifically using the MOD17A3HGF dataset product from 2012 to 2020 (of which 2012–2014 serve only for model spin-up and anomaly baselining; all analyses focus on 2015–2020), with a temporal resolution of 1 year and a spatial resolution of 500 m × 500 m. Temperature and precipitation data were sourced from the CRU_TS meteorological data provided by NCAS (https://crudata.uea.ac.uk/cru/data/hrg/, accessed on 12 August 2023), which offers monthly data covering the land surface from 2012 to 2020 at a resolution of 0.5° × 0.5° (2012–2014 are used exclusively for spin-up, whereas the study window is 2015–2020).

2.2.3. Land Use Data

Land use data were obtained from the MODIS data on GEE (https://modis.gsfc.nasa.gov/data/dataprod/mod12.php, accessed on 12 August 2023), specifically the MCD12Q1 dataset for the period 2012–2020 (data from 2012–2014 support anomaly normalization only; all land- use analyses discuss 2015–2020), with a temporal resolution of 1 year and a spatial resolution of 500 m × 500 m. This dataset classifies land use into six categories: cultivated land, forest land, grassland, water, construction land, and unused land.

2.2.4. Other Data

The DEM data were provided by the Geospatial Data Cloud (https://www.gscloud.cn, accessed on 12 August 2023), managed by the Computer Network Information Center of the Chinese Academy of Sciences, with a temporal resolution of 1 year and a spatial resolution of 30 m × 30 m. Additional data were obtained from the official website of the People’s Government of Heilongjiang Province (https://www.hlj.gov.cn/, accessed on 12 August 2023).

2.2.5. Data Processing

First, the MOD17A3HGF data were stitched and reprojected using the MODIS Projection Tool, and the data were converted from HDF format to TIFF format. The MOD17A3HGF dataset includes quality control files that can be used to verify the quality of the NPP data. All the input data features were normalized using z-score standardization. Missing values flagged in the MOD17A3HGF quality control files were excluded. Outliers were identified using the interquartile range (IQR) method and replaced with the median of neighboring values. This rigorous preprocessing ensured that the input data were clean and suitable for model training. The quality pass rate of the NPP data in this study is as high as 99.7%, indicating that the NPP data are of high quality and suitable for research in the Sanjiang Plain. Next, based on the vector boundary data of the Sanjiang Plain, the NPP data for the study area were clipped using ArcGIS 10.2. Subsequently, the meteorological data, land use data, and terrain data were stitched, reprojected, resampled, and clipped in ArcGIS 10.2 to ensure consistency with the projection, spatial resolution, and extent of the NPP data, thereby addressing the issue of boundary mismatch.

2.2.6. Data Discussion

In this study, the 0.5° × 0.5° resolution CRU_TS meteorological dataset from NCAS was used despite the higher spatial resolution of other datasets such as MODIS MOD17A3HGF and DEM (500 m and 30 m, respectively). This choice is based on the following considerations: meteorological variables like temperature and precipitation generally vary smoothly over space and reflect regional climate trends rather than micro-scale local fluctuations. High-resolution meteorological data often lack long-term continuity or have lower data quality. To ensure spatial and temporal consistency across datasets, the data were reprojected and resampled accordingly. The resulting difference in spatial resolution aligns well with the study’s spatial analysis scale and research objectives, balancing data quality, coverage, and representativeness.

2.3. Methods

2.3.1. NPP-Based Evaluation of Plant Carbon Sequestration Potential

When estimating the carbon source/sink of plants in an ecosystem [37], NEP represents the difference between the soil microbial metabolism and NPP. The formula for calculating NEP is as follows:

$$\mathit{NEP}=\mathit{NPP}-R_H$$

(1)

Using Pei’s model [38], the equation for calculating soil microbial respiration $R_H$ (gCm⁻²a⁻¹) can be derived. The $R_H$-calculation formula is as follows:

$$R_H=0.22\times \left(e^{\left(0.0912T\right)}+\mathit{ln}\left(0.3415R+1\right)\right)\times 3\times 46.5\%$$

(2)

where $R$ is the average annual precipitation (mm) and $T$ is the average annual temperature (°C). A carbon sink occurs when NEP is greater than 0, indicating that the plant ecosystem fixes more carbon than the soil emits. In contrast, a carbon source occurs when NEP is less than 0.

2.3.2. Investigation of Trends

Mann—Kendall trend analysis is a rank-based, non-parametric diagnostic method [39]. The Mann—Kendall trend test was employed in this study to examine the trajectory of NPP from 2015 to 2020. The formula for calculating the Mann—Kendall test result is as follows:

$$S=\sum _{e=1}^{n-1}\sum _{j=e+1}^{n}sgn(X_j-X_e)$$

(3)

where $n$ reflects the dataset’s length, $X_j$ and $X_e$ are information elements corresponding to the $j$ and $e$ time series, and $\mathrm{s}\mathrm{g}\mathrm{n}(X_j-X_e)$ is the sign function, as shown below:

$$\mathit{sgn}\left(X_j-X_e\right)=\left\{\begin{array}{c}+1\left(X_j-X_e\right)>0\\ 0 \left(X_j-X_e\right)=0\\ -1\left(X_j-X_e\right)<0\end{array}\right.$$

(4)

The formula for calculating variance is as follows:

$$\mathit{Var}(S)={\displaystyle \frac{n(n-1)(2n+5)-{\sum }_{j=1}^e{t}_j\left(t_{j-1}\right)(2t_j+5)}{18}}$$

(5)

where $n$ is the length of the dataset, (in this study $n=6$), $e$ is the quantity of groupings, and $t_j$ is the number of value pairs there are in the $j$ category. Then, the common test statistic $Z$ is calculated using the formula:

$$Z=\left\{\begin{array}{c}\left(S-1\right)/\sqrt{\mathit{Var}\left(S\right)} S>0\\ 0 S=0\\ \left(S+1\right)/\sqrt{\mathit{Var}\left(S\right)} S<0\end{array}\right.$$

(6)

where $Z>0$ indicates an upward trend and $Z<0$ indicates a downward trend.

2.3.3. Analysis of Differences

The difference in NPP between 2015 and 2020 at the pixel level was calculated to determine the amplitude of NPP in Guangxi [38,39]. The calculation formula is as follows:

$$D_{ij}=NP{P}_{ij}^{w_1}-NP{P}_{ij}^{w_2}$$

(7)

where $D_{ij}$ is the difference in NPP between the $i$ row and $j$ column pixels in 2015 and 2020; $w_1$ represents 2015; $w_2$ represents 2020; $NP{P}_{ij}^{w_1}$ is the NPP value of the $i$ row and $j$ column pixels in 2015, and $NP{P}_{ij}^{w_2}$ is the NPP value of the $i$ row and $j$ column pixels in 2020.

2.3.4. Analysis of Variational Stability

The trend of NPP fluctuates to some extent due to the combined effects of both natural and human factors. By calculating the coefficient of variation, this study statistically assessed the degree of NPP fluctuation in Guangxi from 2015 to 2020 [40]. The larger the coefficient of variation, the more unstable the changes in NPP. The calculation formula is as follows:

$$B_V={\displaystyle \frac{1}{x}}\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(x_i-\overline{x})}^2}{n-1}}}$$

(8)

$B_V$ is the variational coefficient of NPP in Guangxi from 2015 to 2020; $n$ is the length of the research period in years, which is 6 in this study; $x_i$ represents the NPP value in the $i$ year; and $\overline{x}$ is the typical NPP during the time period under review. The stability degree of variation may be divided into 4 stages based on the computed coefficient of variation [41]: (1) $B_V$ ≤ 0.1, very reliable; (2) 0.1 < $B_V$ ≤ 0.2, reliable; (3) 0.2 < $B_V$ ≤ 0.3, unreliable; (4) $B_V$ > 0.3, very unreliable.

2.3.5. Matrix of Land Use Changes

The relationship between human activities and the natural environment is reflected in land use changes, which can influence NPP as well as the organization and functioning of ecosystems. The land use change matrix can be used to determine the direction, structure, and extent of land use shifts [42]. The following formula is used to calculate the transition matrix:

$$P_{ij}=\left[\begin{array}{ccc}{P}_{11}& \dots & P_{1n}\\ ⋮& \ddots & ⋮\\ P_{n1}& \dots & P_{nn}\end{array}\right]$$

(9)

where $p$ is the location, $P_{ij}$ is the location of the utilization of land classification $i$ converted to land use type $j$, $n$ is the total of the different forms of land usage, and $i$ and $j$ are the two different sorts of land usage, prior to and following its transaction.

2.3.6. Analysis of Limiting Factors

To reduce the impact of additional factors, limiting connection dissection was used in this study to investigate the relationship between NPP and specific weather-related variables [43]. The following formula was used to obtain the partial correlation coefficient:

$$R_{xy,z}={\displaystyle \frac{R_{xy}-R_{xz}{R}_{yz}}{\sqrt{(1-R_{xz}^2)}\sqrt{(1-R_{yz}^2)}}}$$

(10)

Here, $R_{xy,z}$ represents the factor’s part coefficient of correlation $x$ and $y$ with $z$ being held constant, $R_{xy}$ is the relationship value between the $x$ and $y$ variables, $R_{xz}$ is the connection between the $x$ and $z$ variables, and $R_{yz}$ is the ratio of the variables $y$ and $z$ to one another. The relationship coefficients can be computed using the following equation:

$$R_{xy}={\displaystyle \frac{{\sum }_{i=1}^n(x_i-\overline{x})(y_i-\overline{y})}{\sqrt{{\sum }_{i=1}^n{(x_i-\overline{x})}^2{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}}$$

(11)

where $i$ depicts a parameter year, i.e., $i=1,2,3\dots 6$; $n$ reflects the duration of the research, which is 6 years in this study; and $\overline{x}$ and $\overline{y }$ reflect the average amount for the parameters $x$ and $y$. The significance of the partial correlation analysis is tested using the $k$-test, as shown in the following formula:

$$k={\displaystyle \frac{R_{xy,z}}{\sqrt{1-R_{xy,z}^2}}}\sqrt{n-m-1}$$

(12)

Here, $R_{xy,z}$ represents the partial correlation coefficient when the variables $x$ and $y$ are controlled for the variable $z$; $n$ represents the sample size; and $m$ represents the number of variables.

2.3.7. Model Construction

Dataset Analysis

The collected data and the corresponding XCO₂ observations were used as inputs for the correlation analysis (CA), and a covariance matrix was constructed using Equations (13) and (14) to calculate the correlation coefficients, as shown below:

$$\mathit{Cov} X,Y=E\left[\left(X-u_x\right)\left(Y-u_y\right)\right]$$

(13)

$$\rho ={\displaystyle \frac{Cov X,Y}{{\beta }_X{\beta }_Y}}$$

(14)

Here, $u$ is the arithmetic mean of the values, $\beta$ is the standard deviation of the values, $\mathit{Cov}(X,Y)$ is the computed covariance matrix, and $\rho$ is the correlation coefficient. The Pearson coefficient was used to analyze the correlation between the characteristic parameters and the correlation heat map of the relationship between the characteristic parameters and the target value. Through CA, parameters with a correlation coefficient >0.3 were selected as input feature parameters, and the selected feature parameters were CO₂, temperature (Tep), precipitation (Pre), and land type, as shown in Figure 2. The target variable C to be predicted is calculated using Equation (15).

$$C=\phi ({CO}_2,\mathit{Tep},\mathit{Pre},\mathit{Land} \mathit{type})$$

(15)

From the figure, it can be observed that most of the relationships between the feature parameters are nonlinear, and there are also nonlinear relationships between the different predicted target values. This suggests that there are correlations between the spatiotemporal variations in carbon sequestration in Guangxi and certain influencing factors.

Transformer Model

This paper proposes a spatiotemporal variation model for carbon sinks and sources based on Transformer architecture. Transformer is a model architecture that avoids recursion and relies entirely on attention mechanisms to capture global dependencies between inputs and outputs [44]. By discarding the RNN structure and adopting a stacked self-attention mechanism as the core of the network, Transformer utilizes a boundary design, making it suitable for constructing a model to calculate carbon sinks and sources. The self-attention mechanism is a key feature of the Transformer model, as illustrated below:

$$\mathit{Attention}\left(Q,K,V\right)=\mathit{softmax}\left({\displaystyle \frac{{QK}^T}{\sqrt{D_k}}}\right)V$$

(16)

In this equation, $Q\in R^{N\times D}k$ represents the query vector, which is the self-guided prompt in the attention mechanism. $K\in R^{M\times D}k \mathrm{and} V\in R^{M\times D}v$ represent the key and value, respectively, and they correspond to each other in the attention mechanism. The Transformer network model can be divided into two parts: the encoding part and the decoding part. The original Transformer model consists of an encoder and a decoder, as illustrated in Figure 3.

The PCADT model consists of 4 Transformer encoder layers, each with 8 multi-head self-attention heads. The hidden dimension in each layer is set to 256. Following the Transformer layers, two fully connected layers (with 256 and 128 neurons respectively) are used with ReLU activation functions. Dropout with a rate of 0.2 is applied after each fully connected layer to prevent overfitting. The model is trained using the RMSProp optimizer with an initial learning rate of 0.001, batch size of 64, and up to 100 epochs. Early stopping with a patience of 10 epochs based on validation loss was implemented to avoid overfitting.

Carbon Sink Prediction Model Using the Transformer Framework

This study proposes a technical approach, as illustrated in Figure 4, to examine the changes in NPP in Guangxi and its relationship with the influencing factors. The main components of the study comprise three parts: (1) the estimation of NPP in Guangxi using deep learning methods, along with a comparison and validation of the estimation results; (2) the characterization of the spatial and temporal distribution of NPP and its carbon sequestration capacity in Guangxi; and (3) the analysis of the factors influencing NPP changes.

The prediction of NPP is a regression problem, and directly using a Deep Transformer model results in longer computation times and a higher risk of overfitting, making it difficult to achieve optimal performance. On the other hand, employing a shallow Transformer model often fails to effectively capture the interactions between factors, leading to a lack of accuracy in the results. The Transformer model was developed to meet the specific needs of this study, and its overall structure is shown in Figure 5.

This study also employs the mean absolute error (MAE) as a loss function to evaluate the error between the true and predicted values, and uses the adaptive learning rate root mean square propagation (RMSProp) optimization algorithm as the model’s optimizer.

$$\mathit{MSE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{(Y_i-{\widehat{Y}}_i)}^2$$

(17)

Here $Y_i$ is the data from real observations, and ${\widehat{ Y}}_i$ is the data predicted by the Transformer model.

Regarding the benchmark design, to contextualize the performance of the proposed PCADT, we benchmarked it against two widely used classical approaches—Random Forest (RF) and extreme-gradient boosting (XGB)—as well as a naïve mean predictor. All models were trained on the same 2015–2019 grid-cell samples and evaluated on the independent 2020 hold-out set (n = 2146). The hyper-parameters for RF and XGB were tuned via a five-fold cross-validated grid search. The model skill was quantified with mean absolute error (MAE), root -mean -square error (RMSE), and the coefficient of determination (R²). The reported values represent the mean ± standard deviation across the five folds, ensuring a like-for-like comparison.

3. Results

3.1. Verification of the Model

In this study, the remote sensing-based NPP estimation results were compared and validated using China FLUX data. The Transformer model simulated NPP from 2015 to 2020. Figure 6 shows the comparison between the measured data and simulated NPP, with an R² of 0.83 (p < 0.01). Previous studies on NPP estimation in the Hexi Corridor reported an R² of 0.75. When compared to these studies, the precision of the NPP simulation in this research is higher. As summarized in

Table 1. Performance of NPP models for the 2020 grid-cell hold-out set.

Model	MAE	RMSE	R2	R2/FLUX	∆MAE vs. XGB (%)
RF	48.4 ± 2.7	63.9 ± 3.0	0.42	-	−5.1
XGBoost	46.1 ± 2.5	61.8 ± 2.9	0.45	-	0
PCADT	40.4 ± 2.2	55.4 ± 2.6	0.56	0.83	−10.7

, PCADT achieved the lowest MAE (40.4 g C m⁻² yr⁻¹) and RMSE on the grid test set. To shed light on why PCADT outperformed classical models, we visualized the layer-wise attention weights and derived SHAP summaries (Figure 7). Both approaches reveal that the network consistently attends to early -wet -season precipitation anomalies.

3.2. Seasonal Fluctuations in Plant NPP

From Figure 8, it can be observed that the NPP in Guangxi showed an upward trend from 2015 to 2020. The annual average NPP in Guangxi ranged from 885.78 to 906.48 (gCm⁻²a⁻¹), with a yearly mean of 898.69 (gCm⁻²a⁻¹). The lowest and highest values occurred in 2015 and 2018, respectively, with a yearly total NPP ranging from 12400.87 to 12795.85 (gCm⁻²a⁻¹) and an average of 12598.36 (gCm⁻²a⁻¹). Here, UF (k) denotes the forward standardized Mann–Kendall statistic calculated cumulatively from year 1 to year k, while UB (k) denotes the backward statistic obtained by reversing the time series. UF > 0 (or <0) signals an upward (or downward) trend up to k; the year where UF and UB first intersect within the 95% significance band (±1.96) is interpreted as the potential abrupt -change point.

The annual average NPP for each city is presented in

Table 2. Each city’s yearly mean NPP (gCm⁻²a⁻¹) in Guangxi from 2015 to 2020.

Year	Baise	Qinzhou	Beihai	Liuzhou	Wuzhou	Chongzuo	Fangchenggang	Nanning	Laibin	Guilin	Hechi	Hezhou	Guigang	Yulin
2015	1002.62	810.00	659.99	796.27	1104.53	935.35	1038.38	813.73	818.48	792.62	926.58	959.39	841.85	901.08
2016	1012.04	844.10	665.76	793.77	1136.76	981.27	1100.39	848.05	840.81	784.84	941.34	977.14	868.0	923.14
2017	991.60	828.28	661.95	840.78	1142.21	959.24	1040.48	846.34	862.38	833.50	952.11	1017.56	878.87	902.57
2018	1002.52	857.48	699.72	781.38	1148.39	975.72	1093.27	847.05	829.51	795.30	922.64	1004.47	885.66	952.74
2019	1021.96	806.70	675.37	751.93	1065.46	950.17	1026.11	802.75	781.99	746.30	886.61	909.82	821.46	880.88
2020	1007.81	803.12	660.23	790.45	1177.97	946.96	1043.48	827.96	841.67	813.27	932.72	1044.79	885.31	915.02
Mean	1006.43	824.95	670.50	792.4	1129.22	958.12	1057.02	830.98	829.14	794.31	927.00	985.53	863.53	912.57

. The most significant growth was observed in Wuzhou, where the annual NPP increased from approximately 1104.38 (gCm⁻²a⁻¹) in 2015 to nearly 1129.12 (gCm⁻²a⁻¹).

3.3. Spatial—Temporal Variations in NPP Across Different Plant Types

The geographic distribution of the annual mean NPP in Guangxi from 2015 to 2020 is shown in the figure above, which also indicates an upward trend in NPP across the region. From 2015 onwards, the areas with higher NPP are located south of Hezhou, throughout Wuzhou, north of Yulin, and southwest of Fangchenggang. In contrast, the areas with lower NPP are found in Guilin, Laibin, Guigang, and Qinzhou; however, these regions have shown a clear upward trend in NPP (as shown in Figure 9a–f). Overall, the geographic distribution of NPP in Guangxi is heterogeneous, with higher annual NPP in the eastern regions compared to the west (as shown in Figure 9g).

Figure 10a displays the changes in annual NPP in Guangxi from 2015 to 2020. The smaller the interannual difference, the smaller the amplitude of NPP. The results indicate that regions with negative differences account for more than 99%, while positive differences are primarily found in the eastern and western areas. In most western regions, the difference does not exceed 100 (gCm⁻²a⁻¹), whereas in the eastern regions, the difference typically ranges between 0 and 200 (gCm⁻²a⁻¹), with Hezhou primarily showing differences between 100 and 200 (gCm⁻²a⁻¹). Overall, the difference in the eastern region is greater than that in the western region. Regarding the north—south differences, the vast majority are negative, particularly in Qinzhou, where the difference is less than 0 (gCm⁻²a⁻¹).

From 2015 to 2020, the variation coefficient of NPP in the eastern region was higher than that in the western region, showing a tendency for growth from west to east (as shown in Figure 10b). The greater the coefficient, the more volatile the NPP fluctuation. In the northern part of Guilin, the coefficient is generally less than 0.1, indicating relatively stable NPP variation. The coefficients in Hechi, Baise, and Chongzuo are all below 0.2, and their fluctuations remain within a stable range. The NPP coefficients in Hezhou and Qinzhou are relatively high, primarily ranging between 0.2 and 0.3, indicating more unstable NPP variation.

3.4. Evaluation of Plant Carbon Sequestration Potential Based on NPP

Measuring NEP using the relationship model with NPP allowed the researchers to assess plant capacity as a carbon sink. Figure 11 illustrates the plant carbon sink capacity across Guangxi from 2015 to 2020, showing an increasing trend over time, with the most significant increase observed in Wuzhou City in the eastern region. Overall, the plant capacity as a carbon sink varied spatiotemporally and was consistent with NPP. The correlation between the plant carbon sink capacity and NPP was significant, as shown in Figure 12.

Table 3. The overall soil carbon sink of plant and annual mean carbon sink ability in Guangxi.

City	Baise	Qinzhou	Beihai	Liuzhou	Wuzhou	Chongzuo	Fangchenggang	Nanning	Laibin	Guilin	Hechi	Hezhou	Guigang	Yulin
Land area (km2)	36,300	10,897	3337	18,596	12,588	17,300	6173	22,100	13,411	27,800	33,500	11,753	10,602	12,800
Ability to sequester the carbon (gCm−2a−1)	599	491	399	471	672	570	629	494	493	473	551	586	514	543
Total carbon sink (TgC)	21.7	5.4	1.3	8.8	8.5	9.9	3.9	10.9	6.6	13.1	18.5	6.9	5.4	6.9

presents the average plant carbon sink capacity and the overall carbon sinks for 14 cities in Guangxi. The plant carbon sink capacity in most cities was around 400–500 gCm⁻²a⁻¹, with Wuzhou City and Fangchenggang City having a carbon sink capacity exceeding 600 gCm⁻²a⁻¹. The areas with plant carbon sink capacities exceeding 550 gCm⁻²a⁻¹ include Baise City, Wuzhou City, Chongzuo City, Fangchenggang City, Hezhou City, and Yulin City, which play a crucial role in the regional carbon sink.

The total carbon sink for each city (Table 3) was calculated by multiplying the plant carbon sink capacity per unit area (gCm⁻²yr⁻¹) by the corresponding land area (km²). The resulting total carbon sink values were converted to teragrams of carbon (Tg C) using the formula:

$$\mathit{Total} \mathit{Carbon} \mathit{Sink} (Tg C)={\displaystyle \frac{\mathit{Carbon} \mathit{Sink} \mathit{Capacity} (g C m^{-2})\times \mathit{Land} \mathit{Area} \left({km}^2\right)\times {10}^6}{{10}^{12}}}$$

(18)

where the land area is converted from square kilometers to square meters.

To quantify the uncertainty associated with these total carbon sink estimates, we considered two main components, as follows.

NPP Estimation Uncertainty: The uncertainty in plant carbon sink capacity arises primarily from the NPP estimation, which is based on the MOD17A3HGF product. According to previous validation studies, the NPP product carries an uncertainty of approximately ±10%. This reflects errors from remote sensing retrievals, atmospheric corrections, and model assumptions.

Land Area Accuracy: The land area data were obtained from official governmental sources with high spatial accuracy. Thus, uncertainties related to land area measurements are considered negligible compared to those of NPP.

Furthermore, we performed statistical significance tests on the carbon sink estimates using confidence intervals derived from the variability of NPP over the 2015–2020 period. These tests confirm that the spatial differences observed in the total carbon sink across the cities are statistically significant at the 95% confidence level. These uncertainty analyses reinforce the reliability of the presented carbon sink estimates and provide an important context for interpreting regional differences in vegetation carbon sequestration potential.

3.5. Factors Influencing Plant NPP

3.5.1. Analysis of the Limiting Relationships Between Plant NPP and Climatic Parameters

From 2015 to 2020, the typical annual rainfall and temperature in Guangxi exhibited a downward trend, with a decrease of 0.3 °C in temperature and 312.9 mm in precipitation from 2015 to 2020. Spatially, both temperature and precipitation showed an increasing trend from west to east, with the eastern region outperforming the western region (as shown in Figure 13).

The distribution pattern of partial correlations between NPP and temperature is shown in Figure 14a. The partial correlation coefficient between NPP and temperature ranges from −0.998994 to 0.998508. Areas with weak associations are primarily found in the western region, such as Baise City, while most areas exhibit a positive correlation. The correlation pattern between rainfall and NPP is shown in Figure 14b. The Pearson correlation coefficient between NPP and rainfall ranges from −0.998534 to 0.998816. Positive correlations are mainly found in the northern and western regions, whereas negative correlations are more prominent in the southern region.

Overall, the partial correlations highlight spatial disparities between NPP and weather-related variables. While temperature and precipitation show only a moderate correlation, the correlation with precipitation is stronger.

To facilitate interpretation, we classified correlation strengths using the following thresholds:

➀

|R| ≥ 0.7 “Strong correlation”

➁

0.3 ≤ |R| < 0.7 “Moderate correlation”

➂

|R| < 0.3 “Weak correlation”

In Figure 13, darker red or blue areas correspond to |$\rho$| values typically above 0.7, indicating a strong positive or negative relationship with NPP, whereas lighter colors represent weaker correlations. These numerical ranges and threshold definitions are provided in the text (rather than directly on the figure) to preserve the figure’s visual clarity. Readers may refer to this description to accurately interpret the color shading in Figure 13.

3.5.2. Impact of Elevation on Plant NPP

As shown in Figure 15, the pattern of NPP varies with elevation, which can be further divided into four stages. The peak NPP occurs in the range of 200–600 m, corresponding roughly to the western region. NPP then decreases in the range of 600–800 m, forming a low peak area, primarily distributed in the central region. Plant NPP increases at altitudes between 1000–1200 m, and subsequently decreases from 1200 to 2000 m, forming another high—low peak area. Finally, when the altitude exceeds 2000 m, particularly in the southeastern region, the natural conditions become more suitable for plant growth, and NPP shows an upward trend with increasing altitude.

3.5.3. Impact of Land Use Changes on Plant NPP

Figure 16 illustrates the trends in annual NPP for various land use types in Guangxi from 2015 to 2020. Overall, NPP increased across all the land use patterns. Furthermore, the relationship between the cumulative average NPP and the different types of land use is as follows: forest land (1026 gCm⁻²a⁻¹) > grassland (914 gCm⁻²a⁻¹) > cultivated land (694 gCm⁻²a⁻¹) > construction land (211 gCm⁻²a⁻¹) > water bodies (144 gCm⁻²a⁻¹) > unused land (136 gCm⁻²a⁻¹).

Due to the existence of various types of land use, NPP varies across different land use forms, and changes in land use types can further impact NPP.

Table 4. Land use transition matrix of Guangxi between 2015 and 2020.

Land Use Type		2020
		Forest Land	Grass Land	Construction Land	Cultivated Land	Water Area	Unused Land
2015	Forest land	36,371	20,695	0	26	1	8
	Grass land	4625	131,711	185	4874	42	117
	Construction land	0	341	2816	118	5	3
	Cultivated land	11	5020	118	28,504	7	2
	Water area	1	40	4	9	1258	51
	Unused land	10	221	1	7	36	668

presents the land use and its directional changes in Guangxi between 2015 and 2020. In 2015 and 2020, 11 km² and 5020 km² of cropland were converted to forest land and grassland, respectively. Since the NPP of grassland and forest land is higher than that of cropland, the aforementioned land use changes could increase NPP by 4593.17 tC.

4. Discussion

4.1. Spatiotemporal Variation in NPP Across Multiple Plant Species

The spatial variation in net primary productivity (NPP) across Guangxi primarily results from the interplay of climatic conditions, human activities, and ecological restoration efforts. Our Pearson Correlation—Deep Transformer (PCADT) model effectively captures these spatial heterogeneities, producing NPP estimates that closely align with observed environmental gradients.

Climatically, the eastern region benefits from relatively higher precipitation and temperature, creating favorable conditions for plant growth. Specifically, between 2017 and 2019, the annual average temperature and precipitation in eastern Guangxi increased by 0.2 °C and 70 mm, respectively, further enhancing vegetation productivity [2,45,46]. The model’s sensitivity analysis underscores precipitation as the dominant factor influencing NPP spatial distribution.

Human interventions, including afforestation, flood control, and ecological restoration projects, have contributed substantially to vegetation improvement, especially in the eastern region [19,31]. These initiatives correspond with recent regional policies aimed at ecological protection and green economic development.

Topography and environmental fragility also play crucial roles in shaping NPP patterns. The eastern and northeastern karst regions, characterized by ecological vulnerability and geological hazards, display pronounced NPP fluctuations despite recent improvements in vegetation health [19,31,47,48]. Conversely, the western and southern regions exhibit more stable environmental conditions, contributing to historically higher but more gradual increases in NPP.

By integrating climatic, topographic, and anthropogenic factors, the PCADT model elucidates the underlying mechanisms of spatial NPP heterogeneity across Guangxi, providing valuable insights for targeted ecological management and carbon sequestration strategies.

4.2. Climate, Elevation, and Land Use Shape NPP in Guangxi

In wet and somewhat humid environments, temperature is the primary factor influencing NPP fluctuations, while precipitation is the key factor in dry, arid, or semi-arid areas [49]. The results support our conclusions, particularly in the dry, arid, or semi-arid regions of eastern Guangxi. Our partial correlation analyses between NPP and climate factors in Guangxi indicate that NPP is more sensitive to precipitation, which is a key factor limiting its increase. However, the partial correlation between NPP and temperature or precipitation was not significant in most areas, suggesting that these factors are not the primary drivers of the NPP increase in Guangxi. These results align with findings reported by Zhao et al. [50]. Li et al. [46] also emphasized that recurrent human interventions have significantly reduced the impact of climate change on NPP, making human activities the primary driver of NPP changes.

Elevation significantly influences the establishment of biological patterns, plant heterogeneity, and navigation [42,51]. By controlling water, thermal conditions, and soil factors, significant impacts on regional NPP patterns can be observed [16,52]. Considering NPP variation with elevation, we classified areas below 1000 m, between 1000 and 2000 m, and above 2000 m as low-, mid-, and high—elevation zones, corresponding to the western, central—southern, and northeastern regions of Guangxi. Furthermore, the NPP values for each elevation category align with the NPP patterns observed in the different regions of Guangxi. Overall, the bimodal NPP pattern observed in this study resembles those reported by Xiong et al. [47]. Higher NPP values are observed in both the low- and high—elevation regions, with lower values in the mid-elevation region. However, the highest NPP values in the low- and high—elevation regions are not uniform. We compared the NPP changes in the western and central—southern regions of Guangxi, as reported in previous studies [45] and found that the latter exhibited relatively smaller changes. Thus, the central—southern region, which is closer to the eastern region, exhibits the highest NPP values at elevations above 1800 m, with particularly high values observed in the higher elevation zones.

Previous research has shown that human land use and land cover changes alter the composition and functioning of ecosystems, affecting not only landscapes but also carbon fluxes in terrestrial habitats [53]. The impact of land use on NPP can be described in two ways. First, distinct natural and biological systems exist across different land use types. Sandro et al. [54] found that among various land use types, forest land occupies the largest area and has the highest capacity to sequester carbon. This finding aligns with our observation that forest land has the highest NPP. Second, the varying levels of NPP across different land use types suggest that changes in land use can also lead to alterations in NPP.

4.3. Relationship Between Plant NPP and Carbon Sequestration Capacity

The results indicate that the temporal and spatial variations in plant carbon sequestration capacity in Guangxi are closely linked to NPP. Prior research [55,56] has examined this connection, suggesting that the intensity of NPP increase and the ecosystem’s carbon turnover time are the two main factors affecting plant carbon sequestration capacity within ecological communities. Enhancing plant growth and regeneration can lead to higher carbon levels being sequestered in the environment. NPP growth serves as an external driving factor that promotes carbon storage within plant ecosystems, and the effective sequestration of this carbon depends on the ecosystem’s carbon cycle. Additionally, Li et al. [57] studied the relationship between NPP and carbon storage in forests. NPP expansion is a significant driver of carbon sequestration evolution in China’s forest ecosystems, and our findings align with theirs. We also assessed the potential for plant carbon sequestration based on NEP. There is a strong correlation between plant NPP and carbon sequestration capacity, as indicated by both NEP and NPP. NPP represents the organic carbon fixed by plants, minus the amount consumed through autotrophic respiration. NEP is a subset of NPP that remains after accounting for autotrophic respiration. Moreover, extensive evidence shows that inter-annual variability in net primary productivity (NPP) is transmitted almost linearly into changes in ecosystem carbon sequestration at regional scales [58]. In semi-arid environments, water scarcity frequently overrides the nominal CO₂-fertilization signal, leading to pronounced modulation of the vegetation-index-to-GPP relationship [59]. Recent advances in remote sensing of water stress—for example, mapping evapotranspiration anomalies or short-wave -infrared water indices from Sentinel-2—now allow the pixel-level attribution of NPP declines to seasonal droughts. Complementing these vegetation-based indicators, Valerio et al. [60,61] fused Sentinel-1 radar and Sentinel-2 optical data to track the filling and drying of small seasonal ponds in semi-arid landscapes, supplying a dynamic surface -water layer that can be integrated with NPP analyses. Therefore, the spatial and temporal patterns, as well as the influencing factors of plant carbon sequestration capacity correspond closely with NPP.

By assessing the seasonal and spatial variations in plant NPP and carbon sequestration potential, we conclude that Guangxi’s ecosystem is progressively improving. By comparing plant NPP and carbon sequestration capacity across different regions, we can better understand the significant impact of natural conditions on NPP and carbon sequestration, as well as the role of human interventions in mitigating natural limitations. In regions with favorable climatic conditions, both NPP and carbon sequestration capacity remain robust, with human efforts being a critical factor influencing these processes. Therefore, to further enhance NPP levels, strengthening environmental governance and mitigating excessive human impacts on the environment are essential. The primary barrier to increasing NPP and carbon sequestration capacity in areas with less favorable ecosystems is the scarcity of water resources. As environmental conditions improve, both NPP and carbon sequestration capacity will accelerate, though fluctuations are likely to remain more pronounced. Continuous environmental restoration, ecosystem-level improvements, and addressing water scarcity are crucial to reducing NPP fluctuations and further increasing both NPP and carbon sequestration capacity. Additionally, these areas exhibit significant potential for both NPP and carbon sequestration, and the achievable environmental benefits warrant further attention.

4.4. Study Limitations and Future Work

Despite the promising results presented in this study, several limitations should be acknowledged.

First, uncertainties exist in the remote sensing-derived NPP data and meteorological inputs used. These uncertainties arise from sensor calibration errors, atmospheric correction algorithms, cloud contamination, and mismatches in spatial and temporal resolution among datasets. Such factors may affect the accuracy of the NPP estimation and subsequent carbon sink calculations.

Second, the PCADT model relies on specific assumptions regarding feature selection, model architecture, and training data representativeness. While these choices were optimized for the current study area, their applicability to other regions or broader scales may require further validation and adaptation.

In addition, some ecological and anthropogenic variables that potentially influence NPP were not incorporated due to data availability constraints. Future studies should consider integrating such factors to provide a more holistic understanding.

Looking forward, we plan to enhance data quality by combining multisource remote sensing products and higher resolution meteorological data. Model improvements will focus on exploring alternative architectures, such as hybrid models or attention mechanisms tailored to spatiotemporal ecological data. Furthermore, expanding the scope to include land management practices, soil properties, and human disturbance factors will help refine the predictive capabilities. These efforts will aim to improve the accuracy, robustness, and applicability of NPP and carbon sink modeling for better ecological management and climate change mitigation strategies.

5. Conclusions

Our six-year remote- sensing analysis indicates a moderate upward trajectory in both annual vegetation NPP and associated carbon -sink capacity across Guangxi between 2015 and 2020, pointing to a generally improving ecological status. Spatial heterogeneity persists, however: NPP and sequestration potential decrease from west to east, and precipitation exerts a markedly stronger control than temperature, emphasizing that water -resource management remains the primary lever for further gains. Elevation-related bimodality and land -use transitions—especially the conversion of cropland to forest—also shape local patterns, underscoring the importance of targeted restoration in montane zones and the continued protection of existing forest land.

While the PCADT outperforms classical benchmarks, its deep architecture still operates partly as a “black box”. Attention -weight visualization and SHAP attribution suggest an ecologically plausible focus on early -wet -season precipitation anomalies, yet a full mechanistic interpretation requires additional surrogate or hybrid modeling. Moreover, the six-year study window limits our ability to capture decadal climate variability; extending the record and integrating ground-based flux towers would improve robustness.

Taken together, our findings provide indicative guidance rather than a turnkey prescription for precision carbon management. Future work should (i) couple PCADT outputs with hydrological and socio-economic data, (ii) explore ensemble or interpretable surrogate models to validate internal feature interactions, and (iii) test policy scenarios under longer climate baselines. Such steps will be essential for translating short-term gains into resilient, evidence-based strategies for sustaining Guangxi’s vegetation productivity and carbon -sequestration capacity.