An Integrated Framework for NDVI and LAI Forecasting with Climate Factors: A Case Study in Oujiang River Basin, Southeast China

Abstract

In the context of increasingly severe climate change, studying the relationship between climate factors and vegetation dynamics is crucial for ecological conservation and sustainable development. This study focuses on the Oujiang River Basin from 1981 to 2022, aiming to quantitatively model the interactions among temperature, precipitation, the NDVI, and the LAI. Addressing the lack of approaches for forecasting high-resolution LAI data and existing LAI data that are usually interpreted from NDVI data, we proposed a two-step inversion framework: first, modeling the response of the NDVI to climate variables; second, predicting the LAI using the NDVI as a mediating variable. By integrating long-term remote sensing datasets (GIMMS and MODIS NDVI) with meteorological data and applying trend analysis, spatial correlation analysis, and clustering techniques (K-Means and Possibilistic C-Means), we identified spatial heterogeneity in vegetation response patterns. The study results showed that (1) climate factors have a distinctly spatially heterogeneous impact on the NDVI and LAI; (2) temperature is identified as the dominant factor in most regions; and (3) the LAI prediction model based on the climate factors NDVI and NDVI–LAI relationships shows good accuracy in the medium-to-high range of the LAI, with an R² value ranging from 0.516 to 0.824. This study provides a scalable approach to improve LAI estimation and monitor vegetation dynamics in complex terrain under changing climate conditions.

1. Introduction

Vegetation change is an integrated response of ecosystems to climate change and human activities [1]. While numerous studies have documented vegetation dynamics across different climate zones, substantial variability remains in the magnitude and spatial–temporal patterns of change. Compared to extreme environments such as deserts, tropical rainforests, and grasslands, vegetation changes in humid regions, particularly in subtropical areas, tend to exhibit relatively moderate fluctuations [2,3]. Nevertheless, vegetation changes play a crucial role in maintaining the stability and sustainable development of regional ecosystems across various climatic zones. Humid regions, characterized by intensive human populations and economic activities, are particularly vulnerable to climate change, making their vegetation dynamics essential for ecological resilience and sustainability [4]. Additionally, vegetation changes in different ecosystems, including arid, temperate, and subtropical regions, are strongly influenced by climatic factors such as temperature, precipitation, and photoperiod [4].

In recent decades, global climate change has significantly affected vegetation growth, with warming temperatures, shifting precipitation patterns, and increasing atmospheric CO₂ concentrations altering plant phenology and productivity [5,6]. Studies have shown that vegetation in temperate and boreal regions has exhibited a greening trend, largely driven by rising temperatures and lengthened growing seasons [7,8]. Conversely, in arid and semi-arid regions, the impacts of climate change on vegetation are often constrained by water availability [9,10]. In China, extensive research indicates that vegetation coverage has increased nationwide due to climate warming and ecological restoration efforts, such as the Grain for Green Program, but regional differences remain significant [11,12].

In subtropical regions, vegetation changes typically exhibit interannual dynamics characterized by a high correlation with temperature, precipitation, and photoperiod [13,14]. These features not only reflect regional climate variability but also reveal the sensitivity of vegetation to environmental change.

For example, studies in China’s subtropical humid regions have reported strong seasonal variations in vegetation activity, with precipitation playing a dominant role in controlling plant growth during the wet season, while temperature exerts a stronger influence during the dry season [15,16]. However, due to the complexity of the humid subtropical zone and its key role in hydrological processes, current studies have not fully elucidated the relationships between vegetation dynamics and climate factors.

The Normalized Difference Vegetation Index (NDVI), as a commonly used remote-sensing indicator, can effectively characterize vegetation cover and growth conditions by measuring the difference in reflectance between red and near-infrared light [17,18,19]. The value of the NDVI is usually closely related to the density and health of vegetation [20,21] and thus is widely used in agriculture, forestry, and ecology. Especially, the NDVI provides high-frequency and large-scale data support in monitoring the spatial distribution of vegetation cover, growth status, and the effects of drought and disease [22,23]. In existing studies, the NDVI is effective in reflecting vegetation dynamics and its response to climate change. For example, Ghebrezgabher et al. (2020) [24] and Xin et al. (2022) [25] analyzed the significant effects of temperature and precipitation on vegetation growth using NDVI data, which provided an important scientific basis for understanding the complex relationship between climate factors and vegetation changes.

Meanwhile, the Leaf Area Index (LAI), as a key parameter describing the canopy structure of vegetation, has been widely used to study plant photosynthesis, transpiration, and energy exchange processes [26]. Although the traditional in situ LAI measurement method has high accuracy, its application range is limited, and its cost is high [27]. The NDVI-based inversion of the LAI has gradually become an important tool for studying vegetation structure and ecological processes due to its high efficiency and wide applicability. However, the relationship between the NDVI and LAI is affected by a variety of factors [28], including differences in vegetation types, variations in the angle of satellite observations, and interference from atmospheric conditions [29].

The effects of climate change on vegetation in the humid zone are characterized by significant spatial and temporal heterogeneity. The effects of climate factors (e.g., temperature and precipitation) on the NDVI vary significantly across time scales and geographic regions [30]. For example, in some regions, precipitation may have a more pronounced contribution to vegetation, while, in other regions, elevated temperatures may be the main driver of vegetation change. To better understand these complex relationships, researchers have used a variety of data analysis methods, including cluster analysis [31,32]. K-means clustering and Possibilistic C-means (PCM) clustering have become important tools for analyzing the relationship between the NDVI and climate factors due to their flexibility and effectiveness in handling ecological data [32]. K-means can identify potential patterns in the data by minimizing the sum of squares of the distances from data points to cluster centers but is more sensitive to the choice of initial cluster centers. In contrast, the PCM method improves the robustness and adaptability of clustering by introducing fuzzy membership and allowing data points to belong to multiple clusters simultaneously [33].

Despite this progress, several gaps remain. First, studies often lack spatial stratification when modeling the NDVI–LAI relationship, overlooking internal ecological variability. Second, few have applied both K-Means and PCM in comparative frameworks for humid subtropical regions, especially in long-term assessments. Third, while NDVI–LAI modeling is common, its integration with climate drivers in a multi-step framework remains underexplored.

Southeastern China, as a typical area in the humid subtropical zone, has a sensitive response of vegetation to climate change and climatic events. The study of vegetation dynamics in this basin not only helps to reveal the driving role of climatic factors on vegetation changes but also provides an important basis for ecological management and optimal resource allocation in the basin. This study addresses these gaps by proposing a two-step modeling framework linking climate factors to the LAI via the NDVI, enhanced through spatial clustering. The main objectives are as follows: This study focuses on (1) analyzing the spatiotemporal characteristics of the NDVI to identify vegetation trends and patterns; (2) examining the spatiotemporal correlations between the NDVI, temperature, and precipitation to quantify climate-driven vegetation changes; (3) delineating forest zones in the Oujiang River Basin using correlation analysis and clustering models for climate-based ecological classification; and (4) assessing the applicability and accuracy of climate–NDVI and NDVI–LAI relationships, and building an LAI inversion model. The central hypothesis is that temperature and precipitation influence vegetation in spatially heterogeneous ways and that incorporating clustering can reduce uncertainty in NDVI–LAI inversion models. The novelty of this research lies in combining clustering-based spatial stratification with a two-step inversion approach to model vegetation response, particularly in a humid subtropical basin. This integrated method provides a scalable tool for improving LAI estimation and for supporting ecological monitoring and climate adaptation strategies.

2. Materials and Methods

2.1. Data Sources and Data Processing

The Oujiang River Basin, located in the southern part of Zhejiang Province, China, is characterized by a humid subtropical climate with annual precipitation ranging from 1200 to 2100 mm and average temperatures between 15 °C and 17 °C. The region features complex topography, dominated by mountainous terrain, dense forest coverage, and a mosaic of agricultural and urban land use. As shown in Figure 1, the basin’s spatial extent and ecological setting reflect its diverse vegetation types and environmental heterogeneity. This ecological complexity makes the basin highly sensitive to climate variability, providing an ideal setting for investigating vegetation–climate interactions.

Figure 2 and Table 1 present the data used in this study. This study mainly used MODIS NDVI data [34] with a spatial resolution of 250 m × 250 m. MODIS NDVI data had a temporal resolution of 16 days, spanning the period from February 2000 to July 2023, while GIMMS NDVI data [35] had a spatial resolution of 1/12° and a temporal resolution of 15 days, spanning the period from July 1981 to December 2015, and GIMMS NDVI data had a spatial resolution of 1/12° and a temporal resolution of 15 days. The GIMMS NDVI data have a spatial resolution of 1/12° and a temporal resolution of 15 days, spanning from July 1981 to December 2015.

In order to study the long-term NDVI data series in the study area, the GIMMS and MODIS NDVI datasets need to be fused. The correlation coefficient between MODIS NDVI data and GIMMS NDVI data in the study area was 0.87, agreeing with similar studies [36]. Later, the GIMMS data were extended using the method based on MODIS data to obtain long-term data spanning from July 1981 to July 2023.

The GLASS 250-meter LAI (https://glass.hku.hk, accessed on 25 April 2024) is the world’s first LAI product with a spatial resolution of 250 m and a temporal resolution of 8 days, spanning the period 2001 to 2021. Derived from MODIS surface reflectance data, it was developed using a bi-directional short-term and long-term memory (Bi-LSTM) model. The pre-processing workflow consisted of batch conversion of a large number of raster image files in HDF format to TIFF format using the spatial data conversion processing system FME, followed by projection, mosaicking, tiling, and cropping. Finally, eight days of data were aggregated into a monthly image dataset using the Maximum Compositing Method (MVC) approach, providing a more manageable time series for analysis [37]. Meteorological data were derived from the Spatial and Temporal Environment Big Data Platform with a spatial resolution of 1 km and a temporal resolution of 1 month, spanning the period from 1901 to 2021. The unit of temperature data is 0.1 °C, and the unit of precipitation data is 0.1 mm. To facilitate more accurate correlation calculations with the NDVI data, the original 1 km resolution data were resampled to a finer spatial resolution of 250 m using the nearest-neighbor allocation method under resampling. This resampling improves the accuracy of the spatial alignment between the meteorological and NDVI datasets, which is important for understanding their relationship at finer scales [38].

Although extensive preprocessing was applied, uncertainties remain in the remote sensing and climate datasets. NDVI and LAI values may be affected by sensor differences, cloud cover, and interpolation errors. The GIMMS–MODIS NDVI fusion, though supported by high correlation (R = 0.87), may introduce inconsistencies near the transition period. Meteorological data were resampled from 1 km to 250 m, which could affect spatial accuracy. To reduce these impacts, we used MVC for NDVI smoothing and selected Bi-LSTM-derived LAI products with improved temporal consistency [39].

The DEM (Digital Elevation Model) data originates from the ASTER GDEM V3 (30 m) dataset (https://lpdaac.usgs.gov/products/astgtmv003/, accessed on 1 July 2022), which was jointly developed by NASA and JAXA.

Land use data was obtained from the Data Centre for Resource and Environmental Sciences of the Chinese Academy of Sciences (http://www.resdc.cn, accessed on 19 November 2022). The data were mainly based on Landsat TM/ETM remote sensing images with a spatial resolution of 1 km. The data were reprojected for five time periods: 1980, 1990, 2000, 2010, and 2020. The data were classified into six classes: arable land, forest land, grassland, water, construction land, and unused land.

Table 1. Data and methodology.

Name		Period	Spatial Resolution	Temporal Resolution	Methods
The Normalized Difference Vegetation Index (NDVI)	MODIS NDVI	February 2000–July 2023	250 m	16 days	MVC (Maximum Viscosity Composition) [40]
	GIMMS NDVI	July 1981–December 2015	1/12°	15 days		Nearest Neighbor Allocation Resampling
The Leaf Area Index (LAI)		May 2000–December 2021	250 m	1-month
The meteorological data	Precipitation	January 1901–December 2021	1 km	1-month		Nearest Neighbor Allocation Resampling [41]
	Temperature			1-month
Digital Elevation Model		2009	30 m
The land use data		1980, 1990, 2000, 2010 and 2020	1 km	10 years	Intersection

Table 1. Data and methodology.

Name		Period	Spatial Resolution	Temporal Resolution	Methods
The Normalized Difference Vegetation Index (NDVI)	MODIS NDVI	February 2000–July 2023	250 m	16 days	MVC (Maximum Viscosity Composition) [40]
	GIMMS NDVI	July 1981–December 2015	1/12°	15 days		Nearest Neighbor Allocation Resampling
The Leaf Area Index (LAI)		May 2000–December 2021	250 m	1-month
The meteorological data	Precipitation	January 1901–December 2021	1 km	1-month		Nearest Neighbor Allocation Resampling [41]
	Temperature			1-month
Digital Elevation Model		2009	30 m
The land use data		1980, 1990, 2000, 2010 and 2020	1 km	10 years	Intersection

2.2. Research Method

2.2.1. Trend and Changing Point

The Trend-Free and Pre-Whitening Mann–Kendall (TFPW-MK) method was applied for trend and changing point detection of NDVI series as well as climate factor series [42,43,44]. Compared to the classical Mann–Kendall (MK) method, the TFPW-MK method reduces the impact of autocorrelation and seasonality by removing trend components and applying pre-whitening, thereby improving detection accuracy, which is important to climate data and climate-driven factors [45,46].

The linear regression model was used to fit the time-series data (e.g., NDVI and climate data), and the slope of the regression model was used to present the direction and the magnitude of the trend.

2.2.2. Driving Factor Zoning

Zoning based on the driving factors influencing the changes in Normalized Difference Vegetation Index (NDVI) within the Oujiang River Basin provides insights into the extent and spatial distribution of these factors’ impacts on NDVI variations. The zoning was based on the partial and multiple correlation analysis [47,48] between the NDVI and the climate factors at each grid.

The multiple correlation coefficient, denoted as $R$, measures the strength of the linear relationship between one dependent variable and two or more independent variables. It was used to assess the overall fit of a regression model. The formula for calculating the multiple correlation coefficient is as follows:

$$R=\sqrt{1-{\displaystyle \frac{{SS}_{res}}{{SS}_{tot}}}}$$

where ${SS}_{res}$ is sum of the squared differences between the observed and predicted values of NDVI; ${SS}_{tot}$ is the sum of the squared differences between the observed values of NDVI and the mean of NDVI.

The t-test is used to determine the partial correlation between single factor (i.e., precipitation or temperature) and NDVI, while the F-test is used to determine the multiple correlation between the two factors and NDVI [49,50,51]. The t-statistic’s significance level of 0.05 was chosen as the critical value for weak co-driving zone.

The F-test for multiple correlation was used to determine whether the overall regression model, which includes multiple independent variables, significantly predicts the dependent variable.

The F-statistic’s significance level of 0.05 was chosen. The detailed zoning rules can be found in

Table 2. Rules of zoning. $t_T$ is the t-statistic of partial correlation between the NDVI and temperature while controlling for precipitation; $t_P$ is the t-statistic of partial correlation between the NDVI and precipitation while controlling for temperature.

Driving Factor(s)	${\mathit{t}}_{\mathit{T}}$	${\mathit{t}}_{\mathit{P}}$	$\mathit{F}$
Strong co-driving	$t_T>t_{\alpha =0.01}$	$t_P>t_{\alpha =0.01}$	$F>F_{\alpha =0.05}$
Weak co-driving	$t_T\le t_{\alpha =0.01}$	$t_P\le t_{\alpha =0.01}$	$F>F_{\alpha =0.05}$
Temperature	$t_T>t_{\alpha =0.01}$	$t_P\le t_{\alpha =0.01}$	$F>F_{\alpha =0.05}$
Precipitation	$t_T\le t_{\alpha =0.01}$	$t_P>t_{\alpha =0.01}$	$F>F_{\alpha =0.05}$
Not detectable			$F\le F_{\alpha =0.05}$

.

2.2.3. Cluster

In this study, K-Means and PCM clustering methods were selected due to their complementary strengths: K-Means is computationally efficient and suitable for identifying distinct groups in large datasets, while PCM introduces fuzzy membership that better reflects ecological uncertainty and mixed vegetation–climate interactions.

The K-means clustering algorithm used in this study is an iterative solution cluster analysis algorithm [52]. The method first divided the data into N groups, randomly selected N data as the initial clustering centers, and assigned all the data to the closest clustering centers. Based on the assigned data, the cluster centers were recalculated, and the above process was repeated until a certain termination condition was met, which can be that no data is reassigned to different clusters or the cluster centers do not change anymore.

Possibilistic C-Means clustering algorithm (PCM) is a classical clustering algorithm that introduces the concept of “possibility” based on the traditional Fuzzy C-Means (FCM) clustering algorithm to better handle the fuzzy relationships between data points and multiple clusters [33]. The PCM algorithm performs clustering by considering the possibility of data points belonging to different clusters, so it can provide more flexible clustering results for data points that are difficult to clearly classify into a specific cluster. The algorithm first randomly initialized the clustering centers and, given the initial value of the possibility weights, calculated the membership of each data point to each cluster based on the current clustering centers and the possibility weights, where customized possibility weights were used instead of the membership matrix in the traditional fuzzy C-mean algorithm by minimizing the objective function to update the clustering centers and the possibility weights, so that the possibility weights of each data point to each cluster can be adjusted appropriately. Weights were adjusted appropriately until the stopping conditions were met, for example, when the maximum number of iterations (100) was reached or the clustering center changes were small.

The core idea of Davies–Bouldin Index ($DBI$) [53] is to calculate the similarity between each cluster and its most similar cluster and then the average of all similarities to measure the overall clustering results. If the similarity between clusters is higher ($DBI$ is high), it also means that the distance between clusters is smaller, then the clustering results will be worse, and vice versa.

$$DBI={\displaystyle \frac{1}{K}}\sum _{i,j=1}^K\underset{i\ne j}{\max }{\displaystyle \frac{s_i+s_j}{d_{ij}}}$$

where $K$ denotes the number of clusters; $s_{i\left(j\right)}$ denotes the average of the distances from all sample points in the $i$th ($j$th) cluster to the center of the cluster, which is also called the intra-cluster diameter; and $d_{ij}$ denotes the distance between the $i$th and $j$th clusters (i.e., the distance between the centers of the two clusters).

The Davies–Bouldin Index was selected due to its interpretability and effectiveness in unsupervised clustering scenarios, especially for ecological datasets where ground truth is unavailable. It provides a balance between intra-cluster compactness and inter-cluster separation, which is crucial for evaluating vegetation–climate spatial partitions [54].

3. Results

3.1. Trend Changes and Spatial and Temporal Distribution of Vegetation NDVI, LAI, and Climate Variables

3.1.1. Characteristics of Temporal Variations in NDVI, LAI, and Climate Variables

To uncover the temporal evolution of vegetation and climate, we analyzed the monthly NDVI, LAI, temperature, and precipitation data from 1980 to 2022 across the Oujiang River Basin, which were classified into annual (Figure 3A) and quarterly (Figure 3B–E) categories (Figure 3). The NDVI displayed a gradual increase during winter, accompanied by a mild decline or stability in other seasons. This trend suggests a possible extension of the growing season due to winter warming, especially for evergreen or semi-evergreen vegetation. In parallel, the LAI exhibited a significant upward trend, with abrupt changes mainly occurring between 2010 and 2013, as detected by the TFPW-MK method. These abrupt changes align temporally with notable temperature anomalies, indicating that shifts in thermal conditions may have acted as triggers for rapid vegetation expansion. The temperature exhibits fluctuations over time, with a discernible warming trend particularly evident during the winter months. The mutation points predominantly emerge between the years 2006 and 2013. This warming may have alleviated low-temperature constraints on vegetation productivity, explaining the synchronized increases in the NDVI and LAI. Conversely, precipitation showed higher interannual variability and a slight downward trend in spring and autumn, potentially limiting water availability during key vegetative periods.

Overall, the convergence of warming trends and vegetation growth—particularly the alignment of mutation points in the LAI and temperature—highlights temperature as a primary climatic driver in recent decades. These temporal associations reinforce the modeling assumption that temperature plays a dominant role in regulating vegetation dynamics in the region.

3.1.2. NDVI and LAI Spatial Variation Characteristics

The monthly NDVI data for each grid in the Oujiang River Basin from 2000 to 2022 was employed to calculate the quarterly and annual rates of change. The results of linear regression are presented in Figure 4, which illustrates the differences in NDVI changes in the Oujiang River Basin across different periods. The most pronounced vegetation increase occurred during autumn, particularly in mid-elevation zones, suggesting favorable climatic and phenological conditions during this period, such as optimal temperature and delayed senescence. In contrast, winter exhibited relatively weak or slightly negative NDVI changes in most areas, indicating limited photosynthetic activity under colder conditions. This suggests that vegetation growth is more active in autumn, whereas winter exhibits the least variation. In spring, the NDVI demonstrated a slight positive shift, aligning with the general trend of vegetation change observed in the annual average. In summer, the NDVI exhibited a more pronounced upward trajectory compared to spring.

The quarterly and annual changes in the LAI in the Oujiang River Basin were calculated using the same method (Figure 5). Overall, the LAI did not exhibit a statistically significant change, but a slight increasing trend was observed. The most notable increase occurred during the autumn season, while changes in winter were relatively minor and more spatially concentrated. The trends in spring and on an annual scale were similar, both showing minor fluctuations with a relatively stable overall trend. In summer, the LAI trend resembled that of spring, with a generally mild upward tendency, and the red area representing an increase was more widespread.

Together, these results highlight that vegetation responses to climate drivers are seasonally differentiated and spatially heterogeneous—key considerations in modeling regional vegetation–climate relationships.

3.1.3. Time-Series Correlation Between NDVI and Climate Variables

To identify the climatic factors driving NDVI variations, we conducted seasonal and annual linear regression analyses between the NDVI and two key variables—temperature and precipitation—during the period 1981–2022 (Figure 6). The results reveal that temperature plays a dominant role, particularly in spring and winter. The R² value reached 0.311 in spring, indicating a moderate explanatory power, while winter also showed a relatively strong correlation (R² = 0.203). This seasonal pattern suggests that vegetation activity in cooler months is especially sensitive to temperature increases, likely due to the alleviation of low-temperature constraints on plant growth.

In contrast, precipitation exhibited consistently weak correlations with the NDVI across all seasons, with R² values remaining below 0.01 in most cases (e.g., R² = 0.0008 for the annual scale). This may be attributed to complex precipitation–soil moisture–vegetation interactions, which are not effectively captured by simple linear models. Notably, the impact of preceding year climatic conditions was also minimal, although the R² value for the preceding year temperature (0.10184) indicated a discernible influence.

It is worth noting that the relationship between the NDVI and climatic variables can be affected by various factors, including vegetation type, soil moisture, and human activities, and may not be strictly linear. Therefore, the linear regression results presented here primarily capture the dominant trend, while potential nonlinear or compound effects warrant further investigation. That is, temporally separating the time series is not an effective way to build the NDVI–climate factor relationship.

3.1.4. Spatial Distribution of Driving Factor of NDVI

The partitioning of climate drivers in the Oujiang River Basin was performed based on predefined driving criteria, revealing significant spatial differentiation across the region, as depicted in Figure 7. Among the four types of drivers, the weakly co-driven category, influenced by both precipitation and temperature, covers the largest area (

Table 3. NDVI change climate driver types as a percentage of the area.

Driving Factor	Annual	Previous Annual	Spring	Summer	Autumn	Winter
Strong co-driving	0.51	0.37	0.16	1.16	0.18	0.11
Weak co-driving	15.93	19.83	9.77	9.38	8.77	35.03
Temperature	27.45	14.63	13.40	4.15	6.05	18.98
Precipitation	2.64	1.68	0.63	3.05	0.48	0.98
Not detectable	53.47	63.49	76.04	82.26	84.51	44.89

). This suggests that, in most areas, NDVI variability arises from interacting factors, and no single factor is strongly limiting. Conversely, the strong co-driven areas are minimal, highlighting the limited instances where precipitation and temperature exert a substantial combined effect. Temperature-driven zones are more widespread, especially evident in spring and winter, aligning with earlier findings that vegetation is highly sensitive to warming during colder months. Precipitation-driven regions are more localized, mainly in the summer and in areas with complex terrain where water availability becomes a critical constraint.

Notably, climate factors from the previous year had minimal influence on NDVI changes, indicating that current-year climatic conditions play a dominant role in shaping NDVI dynamics, with no significant lag effect observed. This spatial structure of climate–vegetation interactions reinforces the necessity of region-specific modeling strategies: areas with dominant or co-dominant drivers may respond more predictably to climatic variations.

Figure 8 shows the density of temperature-driven grids and precipitation-driven grids of the annual and seasonal NDVI. The co-driven grids are considered with both temperature and precipitation as driving factors, and the grids are used in the density calculation twice. As illustrated by Figure 8, the spatial distribution of driving factors varies for the annual and seasonal NDVI. The overlapping and interlacing of the temperature driving region and precipitation driving region need be taken into account.

3.2. Machine Learning-Based Analysis of the Drivers of Vegetation NDVI Change

3.2.1. Relating NDVI to Temperature and Precipitation Using Different Input Variables

To overcome the limitations of global linear correlations and account for regional variability in vegetation–climate relationships, we applied clustering methods to partition the Oujiang River Basin into climate-responsive zones. Clustering was based on geographic variables including elevation, slope, slope direction, and spatial coordinates (latitude and longitude). These features not only reflect the changes in topography but also help to identify the climatic characteristics of different regions, thus providing strong support for clustering analysis. Clustering also reduces the influence of soil and other factors as well as the uncertainty of NDVI and LAI data [55,56].

The correlation between the NDVI and climatic variables (temperature and precipitation) was analyzed on the basis of zoning by K-Means and PCM clustering methods, and the optimal number of clustering classes was determined by Davies–Bouldin index (DBI). The results (Figure 9) showed that K-Means performed best when the number of clusters was small (2–4), yielding compact and well-separated classes with low DBI values and clear spatial patterns. In contrast, PCM clustering showed sensitivity to the fuzziness coefficient: when the coefficient was set to 2, the clustering quality and climate correlation improved significantly; however, increasing it to 5 or 10 resulted in blurred boundaries and declining differentiation, as reflected by rising DBI values.

Different combinations of input variables have different effects on the clustering effect and correlation. Under the conditions of elevation and slope direction (the first set of input variables), the optimal number of clustering classes mainly focuses on two classes, the clustering effect is the best, and the correlation between NDVI and temperature is the strongest at this time, which indicates that the temperature is the main driving factor affecting the changes in the NDVI. Adding slope as a third variable (second group) increased the optimal number of clusters to three, suggesting that slope introduces additional heterogeneity in terrain-related climate responses. This also modestly enhanced the correlation with precipitation, likely by accounting for slope-mediated moisture redistribution. When latitude and longitude (the third set of input variables) were further added, the correlation between precipitation and the NDVI increased significantly although the DBI values fluctuated, indicating that the spatial location information could better reflect the spatial distribution characteristics of precipitation.

Overall, temperature was the main factor driving the changes in the NDVI (Figure 9A,D,G,J). Slope indirectly enhanced the precipitation correlation by affecting the moisture distribution; latitude and longitude, as spatial location information, further enhanced the ability of the clustering results to capture the precipitation–vegetation response relationship. K-Means was most effective with a smaller number of classes, while PCM clustering was able to effectively characterize the data features with a smaller fuzzy coefficient, enhancing the accuracy of the clustering results.

3.2.2. Comparing Results of Different Methods for Identifying Regional NDVI–Climate Relationships

To assess the consistency of regional NDVI–climate relationships, we compared the spatial distribution of correlation coefficients under different methods, focusing on forested areas with stable land use between 2001 and 2021 to reveal the degree of influence and spatial heterogeneity characteristics of climate factors on vegetation growth (Figure 10). The results showed that the correlation between the NDVI and temperature (Figure 10A) exhibited a high correlation in overall space compared with precipitation, with a higher correlation along the river and in parts of the lower terrain, indicating that temperature fluctuations in these areas have a significant regulatory effect on vegetation growth. The correlation between the NDVI and precipitation (Figure 10B), on the other hand, showed a more dispersed trend, with fewer areas of high correlation and scattered distribution, which may be related to the uneven distribution of precipitation within the watershed and the localized moisture conditions. Overall, the effect of temperature on the NDVI was stronger than that of precipitation, but both had some driving effect on vegetation growth in localized areas (Figure 11).

Figure 12 shows the areas where the correlation coefficient between climate factors and the NDVI is greater than 0.2 (high correlation, see Figure 10) under different clustering methods, including K-Means and PCM (with fuzziness coefficients of 2, 5, and 10). The orders of reach are also illustrated in Figure 12. The results indicate that the high correlation areas are mainly concentrated around rivers. This convergence suggests that these areas exhibit stable vegetation responses to climatic drivers and are thus robustly classified regardless of the method used. From the clustering results, K-Means clustering and PCM clustering (with a fuzziness coefficient of 2) show a high degree of consistency, with good spatial overlap in the distribution of high correlation areas. Figure 13 shows the similarity of the clustering results of different methods. Notably, both K-Means and PCM clustering with a fuzziness coefficient of 2 consistently identified similar zones of a strong NDVI–climate response. The spatial agreement between these methods indicates that certain regions have stable and predictable vegetation–climate relationships. Among the PCM configurations, a fuzziness coefficient of 2 produced the most coherent and interpretable results. Under this setting, class boundaries were clear, and intra-class environmental characteristics were more homogeneous, which in turn enhanced the correlation strength between the NDVI and temperature.

3.2.3. Correlation Analysis Vs. Clustering Models for Climate-Based Zoning

Figure 14 compares the spatial distribution of each classification and driving factors under PCM clustering (with a fuzziness coefficient of 2). For Class 7, precipitation becomes the dominant driving factor, distributed in areas with a high slope (>30°) and high elevation (>1100 m). Vegetation in these areas is highly dependent on moisture, and changes in precipitation directly determine their growth conditions. Class 13 is evenly distributed in areas driven by precipitation and temperature, indicating that both temperature and precipitation have a significant effect. This class is mainly located in the southwestern part of the basin and in areas with moderate slope and high elevation.

In all classes except for Class 7 and 13, temperature is the dominant driving factor. Vegetation in these areas is highly sensitive to temperature changes, especially during the spring and summer when temperature increases lead to accelerated growth. This spatial heterogeneity reflects the differences in responses to climatic drivers among various regions within the basin.

Figure 15 and Figure 16 further support these interpretations by quantifying slope and elevation characteristics for each class and visualizing their respective NDVI–climate correlation strength.

3.3. NDVI and Climatic Variable-Based Inversion Methods for LAI

To quantify the influence of climate on vegetation structure, we extended the NDVI–climate relationship into a two-step inversion framework for estimating the Leaf Area Index (LAI).

To explore the impact of climate factors on the NDVI and the clustering selection results from the previous section, this study employed PCM clustering (with a fuzziness coefficient of 2) and analyzed the relationship between these climate factors (precipitation and temperature) and the NDVI based on annual average data from 2000 to 2021. As shown in Figure 17 and Figure 18, the effects of precipitation and temperature on the NDVI exhibit significant differences across different classes.

The correlation between precipitation and the NDVI is generally weak. Significant correlations between precipitation and the NDVI are only observed in Class 7 and Class 13, with p-values of 0.0322 and 0.0220, respectively. In contrast, the relationship between temperature and the NDVI is stronger. The only classes where temperature is not significantly correlated with the NDVI are Class 7 and Class 13. This indicates that [57] linear regression models can be used to describe the relationships between the NDVI and meteorological data [58], despite their typically nonlinear nature, with manageable uncertainty.

Figure 19 presents a box plot, where vertical lines indicate the distribution range of the LAI within specific NDVI intervals. A significant exponential relationship between the NDVI and LAI was observed across different categories, where the NDVI increased slowly with the LAI at low values and more rapidly at higher vegetation densities. Across all classes, the exponential fit yielded high R² values, confirming the suitability of the NDVI as a proxy for the LAI in spatially consistent ecological zones.

To construct a complete inversion model from climate to the LAI, we combined the linear relationship between temperature and the NDVI presented in Figure 18 and the exponential relationship between the LAI and NDVI shown in Figure 19. A model was developed to estimate the LAI from temperature (Figure 20). The predicted LAI values were compared with observed measurements to evaluate model performance. The model performance under different classes is illustrated in Figure 20, where vertical lines represent 50% of the total data. Overall, the prediction of the LAI performs better in the medium and high ranges, with R² values generally above 0.35 in those regions. However, in low-LAI areas, the model tended to overestimate values. In Classes 1, 4, 5, 6, and 9, the increase in the NDVI and LAI with rising temperature is more pronounced, indicating that suitable temperatures promote vegetation growth. This aligns with prior results showing temperature as the dominant driver in these categories, contributing to better model predictability.

To ensure the reliability of the two-step model, different cross-validation strategies were employed based on the sample size characteristics of each modeling stage. For the temperature–NDVI model, leave-one-out cross-validation (LOOCV) was applied, with the data consisting of the annual mean temperature and NDVI samples during the study period. Given the limited number of samples in this model, LOOCV provided a relatively objective evaluation of model performance under small-sample conditions. For the NDVI–LAI model, five-fold cross-validation (5-fold CV) was employed, using grid-based monthly observational data from 2001 to 2021, which contained a large sample size. Within each class, the dataset was randomly divided into five subsets; in each iteration, four subsets were used for model training, and the remaining subset was used for validation, repeated over five cycles. Unlike the conventional approach of averaging the evaluation metrics across folds, this study aggregated all validation predictions within each class across the five folds and calculated the coefficient of determination (R²) and root mean square error (RMSE) based on the overall merged predictions and true values. This approach captures the full prediction variance and better reflects model generalization ability. As summarized in

Table 4. Model performance metrics (R² and RMSE) for each class based on cross-validation.

Class	R2	RMSE	Class	R2	RMSE
Class 1	0.374	0.390	Class 8	0.317	0.366
Class 2	0.345	0.634	Class 9	0.345	0.434
Class 3	0.373	0.565	Class 10	0.318	0.379
Class 4	0.360	0.457	Class 11	0.338	0.372
Class 5	0.370	0.302	Class 12	0.370	0.438
Class 6	0.373	0.306	Class 13	0.266	0.328
Class 7	0.289	0.402	-	-	-

, R² values across classes ranged from 0.266 to 0.374, while RMSE values ranged from 0.302 to 0.634. Although these metrics showed a slight decline compared to the in-sample fitting, the consistency across categories and the limited error increase underscore the robustness and practical applicability of the proposed climate-driven LAI inversion model.

4. Discussions

The observed vegetation dynamics in the Oujiang River Basin align with typical warming trends in humid subtropical climates. The observed strong correlation between the NDVI and temperature during spring and winter suggests that thermal conditions are a key driver of early vegetation activity. In contrast, the relatively weak correlation with precipitation implies that water availability is generally not a limiting factor.

This study employed a two-stage regression framework, first establishing the relationship between temperature and the NDVI and subsequently using the NDVI to estimate the LAI. The use of the NDVI as an intermediate variable allows the separation of direct and indirect climate influences. Exponential models were applied to characterize the NDVI–LAI relationship, capturing nonlinear saturation patterns observed in areas with dense vegetation cover [59]. This modeling strategy is consistent with prior research that demonstrated the effectiveness of exponential functions in improving LAI estimation performance [60]. In our analysis, this approach yielded favorable results, particularly in regions with moderate-to-high vegetation density.

The primary objective of LAI prediction in this context is to support ecohydrological and ecosystem process modeling. Annual LAI estimates serve as key inputs to models such as SWAT and Biome-BGC, which simulate vegetation growth, evapotranspiration, and carbon fluxes [61]. By combining the NDVI and climate data from the current year, the model enables projection of the LAI for the following year, offering practical value for forest monitoring, watershed planning, and ecological forecasting. The method is particularly suitable for natural forested landscapes with minimal anthropogenic interference. However, in areas affected by frequent land use changes or human activity, the predictive capacity of climate-based models diminishes due to the influence of non-climatic factors.

While the modeling results were validated using LOOCV and 5-fold cross-validation, several sources of uncertainty remain. These include potential noise in remote sensing data, artifacts introduced during NDVI–LAI integration, and spatial mismatch arising from the resampling of meteorological variables. The clustering methods used—K-Means and PCM—were effective in delineating spatial patterns but may oversimplify ecological gradients or transition zones. Additionally, while the Davies–Bouldin Index was employed to assess clustering performance, further evaluation using complementary metrics such as the Calinski–Harabasz Index or silhouette scores could strengthen the analysis.

The broader applicability of this modeling framework requires further investigation. The relationship between vegetation indices and climatic variables may differ across forest types, ecological zones, and climatic regimes. Future research should evaluate model transferability across diverse environments, incorporate vegetation-type-specific calibration, and consider lagged or cumulative climate metrics such as seasonal rainfall totals and heat accumulation. Integrating additional environmental variables—such as soil moisture and solar radiation—and cross-validating results with other LAI products would further improve the ecological robustness and predictive reliability of the model.

5. Conclusions

This study examines the influence of climate factors on vegetation dynamics in the Oujiang River Basin from 1981 to 2022, utilizing NDVI, LAI, temperature, and precipitation data, complemented by clustering analysis to assess spatial heterogeneity. The key findings are as follows:

(1)

Temperature emerges as the primary driver of the NDVI, particularly in spring and winter, while precipitation exhibits a weaker influence. This is consistent with findings from similar studies in other humid subtropical regions, where temperature has been shown to significantly influence seasonal vegetation growth [62].

(2)

Analysis of the drivers for each grid revealed that temperature-driven effects dominate in low-elevation zones, while precipitation-driven effects are concentrated in areas with a high slope (>30°) and high elevation (>1100 m), highlighting spatial variability in climatic drivers. However, treating the entire watershed as a whole or analyzing each grid individually will lead to non-negligible uncertainties.

(3)

The machine learning classification method can categorize grids into a certain number of classes, maximizing the potential for establishing NDVI-meteorological data relationships. This classification approach helps reduce modeling uncertainty and enhances the spatial consistency of climate–vegetation analysis. For all classes, the NDVI and LAI demonstrate a significant logarithmic relationship, with R² values exceeding 0.90 across most clustering categories. The LAI prediction model, based on temperature–NDVI and NDVI–LAI relationships, performs effectively for medium-to-high LAI values.

In summary, vegetation dynamics in the Oujiang River Basin are primarily driven by temperature in low-elevation areas and precipitation in high-elevation (elevation > 1100 m) zones, with the NDVI and LAI showing a robust correlation and effective LAI prediction for mid-to-high values. These findings validate the integrated two-step modeling framework and offer a potential tool for vegetation monitoring and forest hydrological assessment in climate-sensitive regions. Future research should integrate additional environmental variables and field observations to reduce uncertainty in LAI forecasting at the grid scale and enhance result reliability and applicability.