From the Past to the Future: Unveiling the Impact of Extreme Climate on Vegetation Dynamics in Northern China Through Historical Trends and Future Projections

Abstract

Over the past few decades, occurrences of extreme climatic events have escalated significantly, with severe repercussions for global ecosystems and socio-economics. northern China (NC), characterized by its complex topography and diverse climatic conditions, represents a typical ecologically vulnerable region where vegetation is highly sensitive to climate change. Therefore, monitoring vegetation dynamics and analyzing the influence of extreme climatic events on vegetation are crucial for ecological conservation efforts in NC. Based on extreme climate indicators and the Normalized Difference Vegetation Index (NDVI), this study employed trend analysis, Ensemble Empirical Mode Decomposition, all subsets regression analysis, and random forest to quantitatively investigate the spatiotemporal variations in historical and projected future NDVI trends in NC, as well as their responses to extreme climatic conditions. The results indicate that from 1982 to 2018, the NDVI in NC generally exhibited a recovery trend, with an average growth rate of 0.003/a and a short-term variation cycle of approximately 3 years. Vegetation restoration across most areas was primarily driven by short-term high temperatures and long-term precipitation patterns. Future projections under different emission scenarios (SSP245 and SSP585) suggest that extreme climate change will continue to follow historical trends. However, increased radiative forcing is expected to constrain both the rate of vegetation growth and its spatial expansion. These findings provide a scientific basis for mitigating the impacts of climate anomalies and improving ecological quality in NC.

1. Introduction

Since the 20th century, there has been a notable increase in the Earth’s surface temperature, resulting in more frequent occurrences of extreme weather events. These include extended periods of drought and heavy rainfall, which have had profound effects on the global environment, economic systems, and public health [1,2,3]. Statistical data indicate that more than 10,000 extreme climate events were recorded worldwide between 1997 and 2016, resulting in economic losses exceeding USD 3 trillion and approximately 524,000 fatalities [4]. Simultaneously, these extreme climatic events have substantially disrupted regional vegetation dynamics and ecosystem service functions [5], thereby attracting widespread public and scientific attention globally [6].

Vegetation plays a critical role in biogeochemical cycles [7], climate regulation, and carbon sequestration [8]. Monitoring vegetation dynamics under global warming is essential for understanding shifts in global ecosystems [9]. Remote sensing technology has been widely applied to monitor vegetation across large areas. The Normalized Difference Vegetation Index (NDVI) provides a precise representation of vegetation coverage at the pixel scale and functions as a critical metric for evaluating long-term and large-area vegetation dynamics. It is extensively utilized to analyze vegetation trends and evaluate their responses to climatic factors [10,11]. Thus far, most studies have focused on the effects of mean-state climate on vegetation [7,12], paying less attention to the impacts of climate extremes on vegetation dynamics. These studies have primarily examined a single timescale [13] or considered multiple timescales assuming linearity and smoothing [14]. In reality, both the NDVI and climate variable time series exhibit non-smooth and nonlinear characteristics, necessitating the use of appropriate methods for decomposing these into distinct timescale components. The ensemble empirical modal decomposition (EEMD) method, extended by empirical mode decomposition (EMD), facilitates the examination of vegetation dynamics’ responses to climate variations across multiple temporal scales [15,16,17]. Furthermore, the existing research has largely concentrated on the impacts of current climate conditions on vegetation, with fewer studies addressing the potential ramifications of future extreme climate changes for vegetation [18,19,20]. Under a global high-carbon emission scenario, the frequency and intensity of future extreme climate events, such as droughts and heat waves, are projected to increase substantially. Consequently, there is an urgent need for high-precision climate prediction data that will allow us to evaluate their ecological impacts.

Northern China (NC), situated in the mid-latitude zone of East Asia, encompasses arid, semi-arid, and semi-humid zones characterized by low precipitation, scarce water resources, low vegetation cover, and fragile ecological resilience [21]. Due to global warming, the climate in NC has undergone substantial changes, particularly in terms of extreme weather conditions that may affect regional vegetation dynamics [22]. For instance, studies have demonstrated that climate is a predominant factor influencing vegetation growth in specific regions, such as the Qilian Mountains [23] and the Loess Plateau [24], with discernible temporal and cumulative effects. Moreover, research indicates that extreme climatic events have significantly affected vegetation dynamics in Northwest China [25], Inner Mongolia [26], and the agricultural–pastoral transitional zone [27]. Compared to other regions, the climatic characteristics of NC render vegetation activity more susceptible to extreme climatic events [27,28]. However, systematic research on the historical and prospective impacts of extreme climate events on the ecologically fragile regions of northern China remains limited, particularly with regard to the quantitative analysis of vegetation–climate relationships across different timescales, the identification of key driving factors, and the projection of future change trends—areas where the current understanding is still insufficient. To address these gaps, this study aims to (1) employ the Sen’s slope and Mann–Kendall methods to investigate the spatiotemporal dynamics of the NDVI and extreme climate indices in northern China (NC); (2) determine the dominant timescale of NDVI variation in NC using Ensemble Empirical Mode Decomposition (EEMD), and quantitatively assess the spatiotemporal response of the NDVI to historical extreme climate indicators at multiple timescales through pixel-wise correlation analysis; (3) identify the key extreme climate variables significantly influencing vegetation in NC using all-subsets regression; and (4) integrate data from four CMIP6 models to project future spatiotemporal trends of key extreme climate indices from 2025 to 2060 and develop a random forest model to evaluate the potential impacts of these key climate variables on the NDVI under various future scenarios. This study offers a valuable reference for understanding vegetation dynamics in regions with similar climatic conditions and for minimizing the detrimental consequences of extreme climatic events on ecological systems.

2. Materials and Methodology

2.1. Study Area

The NC (73°37′–125°57′ E, 31°41′–53°18′ N) encompasses 15 provinces and municipalities: Jilin, Liaoning, Heilongjiang, Inner Mongolia, Hebei, Tianjin, Beijing, Shandong, Henan, Shanxi, Shaanxi, Gansu, Qinghai, Ningxia, and Xinjiang (Figure 1). It spans an approximate area of 5.62 × 10⁶ km², accounting for 58% of China’s total land area. The terrain progressively rises from the northeast to the southwest, with most elevations ranging between 500 and 2000 m. The climate predominantly follows a temperate-continental along with a temperate-monsoon climate pattern. Winter temperatures drop below 0 °C, while summer temperatures exceed 20 °C. Accordingly, the lowest and highest monthly average temperatures occur in January and July, respectively. Annual precipitation remains low and predominantly occurs during the summer season [29]. The NC boasts abundant vegetation, yet its ecosystem remains highly vulnerable. Due to its extensive longitudinal span, the study area exhibits distinct horizontal vegetation zonation influenced by climatic factors. From east to west, natural vegetation zones transition from forests and grasslands to typical grasslands, meadows, desert steppes, and ultimately deserts.

2.2. Data Sources

The monthly NDVI dataset for China, featuring a spatial resolution of 5 km × 5 km from 1982 to 2018, was generated by monthly maximum synthesis, combining and extracting data from the NASA AVHRR CDR NDVI V5 dataset (http://www.geodata.cn (accessed on 30 August 2024)). It comprises 468 images in TIF format, representing the monthly synthesized NDVI. In this article, a time series dataset of the NDVI for NC was developed by implementing mask extraction, reprojection and averaging techniques. Additionally, areas of bare land characterized by an annual average NDVI below 0.1 were excluded to minimize potential bias.

The climatic data was sourced from the National Meteorological Information Center (http://data.cma.cn/ (accessed on 25 July 2024)). Three key variables (daily maximum temperature, daily minimum temperature, and daily precipitation) were extracted from 415 ground meteorological stations in NC for the period 1982–2018. In cases where missing data were identified, meteorological records from adjacent time points were utilized for data imputation. Following the framework of the 27 climate indices defined in the IPCC Fourth Assessment Report, we selected four extreme precipitation indicators and five extreme temperature indicators that are particularly relevant to droughts, floods, high-/low-temperature events, and vegetation growth. All indices were calculated using the original daily precipitation and temperature data. Finally, a high-resolution (5 km × 5 km) gridded dataset was generated by applying the Kriging method to the station-based extreme climate indices.

The climate simulation data was sourced from the 10 km HiCPC (High-Resolution CMIP6 China Daily Downscaled Climate Prediction) dataset provided by the National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/ (accessed on 2 December 2024)). This dataset employed the advanced Bias Correction and Spatial Disaggregation (BCSD) method to address the inherent biases in GCM simulations [30]. Under two different scenarios—SSP245 and SSP585—climate data from the NESM3, INM-CM5-0, MRI-ESM2-0, and ACCESS-CM2 models for the period 2025–2060 were utilized to compute future extreme climate indices based on daily precipitation, daily maximum temperature, and daily minimum temperature data in this research. Additionally, the dataset was resampled to a spatial resolution of 5 km for subsequent analysis.

2.3. Methodology

2.3.1. Calculation of Extreme Climate Indices

In this study, the RClimDex 1.0 software developed by Canadian Center for Meteorological Research (CCMR) was utilized to compute all extreme climate indices [31]. It takes daily maximum temperature, minimum temperature and precipitation data as input. We selected 9 extreme climate indicators to participate in the calculation (

Table 1. Definitions of extreme climate indices.

	ID	Name	Definition	Unit
Extreme temperature indices	TNn	Min Tmin	Monthly minimum value of daily minimum temp	°C
	TXx	Max Tmax	Monthly maximum value of daily maximum temp	°C
	TN10P	Cool nights	Percentage of days when Tmin < 10th percentile	d
	TX90P	Warm days	Percentage of days when Tmax > 90th percentile	d
	DTR	Diurnal temperature range	Monthly mean value of difference between Tmax and Tmin	°C
Extreme precipitation indices	RX1D	Max one-day	Monthly maximum consecutive 1-day precipitation	mm
	RX5D	Max five-day	Monthly maximum consecutive 5-day precipitation	mm
	CDDs	Consecutive dry days	Maximum number of consecutive days with RR < 1 mm	d
	CWDs	Consecutive wet days	Maximum number of consecutive days with RR ≥ 1 mm	d

), resulting in a comprehensive dataset encompassing extreme temperature and precipitation events from 1982 to 2018.

2.3.2. Changing Trend

The Theil–Sen method is a nonparametric technique for estimating trends which is known for its robustness and computational efficiency, and thus, it is widely employed in analyzing long-time series data trends [32]. The sign of the slope indicates an increasing or decreasing trend. However, due to the absence of a statistical significance test for trends in this method, we employ the Mann–Kendall test to evaluate it.

The Mann–Kendall test represents a nonparametric statistical approach that has the advantage of not assuming normal distribution or linear trends for independent variables. It is also unaffected by missing or outlier values. In this study, the above two methods are utilized to analyze NDVI change trends [33]. In this research, the significance threshold is set at α = 0.05, with Z_1−α/2 having a value of 1.96 (where α refers to the level of statistical significance).

2.3.3. Ensemble Empirical Mode Decomposition

This research investigates the spatiotemporal variations in the NDVI by applying the Ensemble Empirical Mode Decomposition (EEMD) method. As a time-frequency analysis approach, Empirical Mode Decomposition (EMD) decomposes a signal into several Intrinsic Mode Functions (IMFs) as well as a residual component (RES) [34]. Despite its utility, EMD faces issues with signal mixing across various modes. To tackle this, the EEMD method has been introduced as an enhancement [15]. The steps of the decomposition process are outlined in detail below:

(1)

A specific quantity (np) of Gaussian white noise, denoted as ω_j(t), is incorporated into the primitive signal x(t), resulting in the generation of numerous new signals that contain noise:

$$X_j(t)=x(t)+{\omega }_j(t)$$

(1)

$$j=1,2,\cdot \cdot \cdot ,np$$

(2)

(2)

X_j(t) is analyzed by using the EMD model to extract the IMF components c_i and the component r_n:

$$X_j(t)={\displaystyle \sum _{i=1}^n{c}_{ij}+r_{nj}}$$

(3)

where c_ij and r_nj represent the ith IMF and the trend term r_n, derived from X_j(t).

(3)

Reiterate the aforementioned steps for np iterations, incorporating a novel white noise sequence of equivalent magnitude during each iteration.

(4)

Merge and average the extracted IMF and residual term r_n from each decomposition to eliminate the aggregate influence of the incorporated white noise.

$$c_i(t)={\displaystyle \frac{1}{np}}{\displaystyle \sum _{j=1}^{np}{c}_{ij}}(t)$$

(4)

$$r_n={\displaystyle \frac{1}{np}}{\displaystyle \sum _{j=1}^{np}{r}_{nj}}(t)$$

(5)

where c_i (t) represents the ith component derived from decomposing the original signal.

The fluctuation patterns of the IMF components are illustrated as follows:

$$T_k={\displaystyle \frac{N}{N{P}_k}}$$

(6)

where the number of data points is N in the IMF component, and NP_k denotes the count of extrema within the kth IMF component.

The importance of the IMF component across various temporal scales may be quantified by its contribution to the variance, as detailed below:

$$S_i={\displaystyle \sum _{i=1}^k{(c_i(t))}^2}-{\left({\displaystyle \sum _{i=1}^k{c}_i(t)}\right)}^2$$

(7)

$$S={\displaystyle \frac{S_i}{{\displaystyle \sum _{i=1}^{i-1}{S}_i}}}\times 100\%$$

(8)

where S and S_i represent the total variance devotion and the devotion of the ith component, respectively.

2.3.4. Correlation Analysis

We assess the influence of different extreme climate indicators on the NDVI across various timescales through image-by-image metrics by utilizing Pearson correlation analysis, with a significance level of 0.05 [35]. The formula for the calculation is presented below:

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

(9)

where R_xy denotes the correlation coefficient between variables. Let x_i and y_i represent the NDVI and extreme climate index (ECI) values of the ith year, respectively, and let $\overline{x}$ denote the mean NDVI across all years. $\overline{y}$ is the mean ECI value across all years and seasons and i represents the sample count. The computations were conducted using MATLAB 2018b [36].

2.3.5. Full Subset Regression

Full subset regression is a statistical technique that is employed to explore and model relationships between variables, aiming to identify the optimal model from all possible combinations of independent variables. The core principle involves exhaustively evaluating all possible subsets of independent variables, assessing the regression model for each subset, and choosing the optimal model according to specific criteria, such as the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), or adjusted R². Unlike heuristic-based methods such as stepwise regression, full subset regression does not depend on the order of variable selection. Instead, it systematically examines all possible variable combinations, ensuring that no potentially important models are overlooked [37]. In this research, full subset regression is utilized to conduct regression analysis between extreme climate variables and the NDVI. The best-fitting model is determined using the adjusted R² value. Additionally, the evaluation of model selection is carried out through the determination coefficient (R²) and the Bayesian Information Criterion (BIC). Specifically, models with higher R² values and lower BIC values are compared and analyzed to determine the optimal combination of variables.

2.3.6. Random Forest

To forecast the potential effects of extreme climate dynamics on NDVI variations in NC in the future, this research utilized machine learning techniques. We conducted a comprehensive analysis of various machine learning techniques such as multiple regression and support vector machine (SVM) used in previous studies. Compared to the support vector machine (SVM), which requires extensive parameter tuning, and neural networks, which are highly dependent on large datasets and prone to overfitting, the random forest (RF) method demonstrates superior performance with relatively minimal parameter adjustment. It offers notable advantages in ecological modeling by effectively managing noisy data, capturing nonlinear relationships between vegetation and extreme climatic factors, and enabling the assessment of feature importance to evaluate the influence of each predictive variable. Based on these merits, we selected the random forest as the primary analytical model for this study. The random forest algorithm, proposed by Breiman [38]), is a supervised learning technique that aggregates the outputs of multiple tree-based models to construct an optimal model for a specific application. This algorithm generates a large number of decision trees by randomly partitioning the training dataset into multiple subsets, thereby achieving superior performance compared to a single decision tree.

The research methodology comprises four sequential steps (Figure 2). (1) Extract extreme climate indicators and the NDVI dataset from designated data sources following a full subset regression screening process to select relevant variables for model input. (2) Perform essential data preprocessing procedures, including anomaly filtering, partitioning the dataset into training and test subsets (80% for training and 20% for testing), and normalization to ensure uniformity in variable scales. The data spanning from 1982 to 2018 were processed accordingly using the createDataPartition function in the caret package of R 4.4.2 software [39]. (3) Construct a predictive model for future NDVI values using the random forest algorithm. To mitigate the risk of overfitting, five-fold cross-validation was implemented. Based on prior research [39], the number of trees was set to 1000, and mtry was determined as 4. (4) Evaluate the model performance using multiple quantitative metrics, such as the coefficient of determination (R²), root mean square error (RMSE), mean squared error (MSE), mean absolute deviation (MAE), and the explained variance metric (EVS), to ensure that the model meets the predefined performance standards and research objectives.

3. Results

3.1. Spatiotemporal Characteristics of Vegetation Dynamics in NC

The mean annual NDVI in NC fluctuated between 0.375 and 0.457, showing an increasing trend of 0.0030/a from 1982 to 2018 (Figure 3). The lowest value was recorded in 1989, while the highest value reached 0.4571 in 2018 over the study period. Spatial analysis of the annual average NDVI revealed an overall improvement in vegetation coverage across NC, albeit with localized degradation trends (Figure 4). Approximately 65.6% of the study area exhibited an increasing trend in the NDVI, of which 50.1% showed a statistically significant increase. These regions were primarily distributed in northwest Xinjiang, eastern Qinghai, the Loess Plateau, and Hebei Province. Conversely, about 32.4% of the total area experienced a decrease in NDVI values, with the most pronounced declines observed in central and western Inner Mongolia, the western Hexi Corridor, and southern Xinjiang.

3.2. Periodicity Analysis of Vegetation Dynamics in NC

The NDVI was decomposed into four Intrinsic Mode Functions (IMFs), with oscillation periods of 3, 6, 12, and 37 years, respectively (Figure 5). Among these, IMF1 contributed the most to variance, accounting for 26.60%, followed by IMF2 at 10.1%. The variance contributions for IMF3, IMF4, and the residual component (RES) were 10.3%, 5.35%, and 24.5%, respectively (

Table 2. Timespans and variance contributions of IMF components within the NDVI series.

	IMF1	IMF2	IMF3	IMF4	Residual (RES)
Periodicity	3 *	6	12	37	— —
Variance Contribution (%)	26.6	10.1	10.3	5.35	24.5

* Passing the 95% significance test.

). Periodicity tests conducted on both IMF1 and RES revealed statistically significant results, passing the significance threshold of p < 0.05 at the 95% confidence level. Notably, the RES demonstrated an upward tendency, suggesting a continuous increase in the NDVI in NC from 1982 to 2018. In summary, the NDVI in NC predominantly displayed cyclic variations occurring every three years along with a long-term increasing trend.

3.3. Spatiotemporal Characteristics of Extreme Climate Indices in NC

It can be observed from Figure 6 that the mean values of DTR (diurnal temperature range), TN10P (cool nights), TNn (Min Tmin), TX90P (Warm days), and TXx (Max Tmax) for all years were 12.9 d, 12.76 d, −24.06 °C, 12.46 d, and 34.76 °C, respectively. DTR and TN10P varied and declined, with rates of −0.0097d/a and −0.2594 °C/a, respectively, while TNn, TXx, and TX90P fluctuated and increased, with rates of 0.0063 °C/a, 0.0503 °C/a, and 0.2530 d/a, respectively. Extreme precipitation indices exhibited a notable upward trend (Figure 7), with the trend for RX5D (max five-day) being more pronounced than that for RX1D (max one-day). Specifically, RX1D and RX5D increased by 0.07 mm/a (p < 0.05) and 0.09 mm/a (p < 0.05), respectively. The CDDs (consecutive dry days) significantly declined at a rate of −0.2545 d/a, and CWDs (consecutive wet days) insignificantly decreased at a rate of −0.0012 d/a. Both CDDs and CWDs demonstrated a shortening trend in the study area, suggesting that the precipitation in NC presented a balanced trend.

Spatially, DTR varied from -0.05 to 0.049, with declines concentrated in the northwest of NC (Figure 8). As can be observed, TN10P exhibited a predominantly decreasing pattern, with a significant reduction evident in western Qinghai. The changing trend of TNn exhibited significant spatial heterogeneity, with most regions showing an upward trend. TX90P had the most obvious rising tendency in the central part of NC. In contrast, TXx presented an increasing trend in almost the entire NC. Changes in CDDs ranged from -1.418 to 0.887, with the most obvious decreases occurring in eastern Xinjiang and northern Gansu, and the most evident increases in central-western Qinghai Province (Figure 9). Areas experiencing significant increases in CWDs were primarily located in western Xinjiang, while noticeable increases were also observed in Jilin. The spatiotemporal changes in RX1D and RX5D were similar, showing increasing trends in the western and central parts of NC, but decreasing trends in northeastern NC, Hebei, and Inner Mongolia.

Figure 10 and Figure 11 show the outcomes of the significance analysis for the spatial variations in each extreme climate indicator. DTR passed the significance test in Heilongjiang, Shandong, Henan, Qinghai, and Xinjiang. TN10P exhibited a notable decline across most regions, while TNn presented a significant increase in Qinghai, Henan, Shanxi, Shandong, and Hebei. TX90P and TXx presented significant increases at the vast majority of sites, with TX90P failed the significance test only at individual sites in Heilongjiang and western Xinjiang. The CDDs significantly decreased at some sites in northern and eastern Xinjiang and northeastern China. The CWDs notably increased in western Xinjiang but decreased in northern and northeastern NC. The significance test results for RX1D and RX5D were largely consistent, showing significant increases at sites in Xinjiang, Qinghai, Gansu, and Ningxia.

3.4. The Relationship of Extreme Climate Indices and Vegetation in NC

To explore the spatiotemporal heterogeneity of the NDVI in NC, a multi-temporal scale analysis was performed at the pixel level. The correlation analysis at the three-year timescale (Figure 12 and

Table 3. Statistics of correlation coefficients between extreme temperature indicators and the NDVI at different timescales (%).

Timescale	Correlation	TN10P	TNn	TX90P	TXx	DTR
3a period	snc	36.80	1.40	8.00	3.60	15.30
	isnc	29.60	32.90	27.40	33.70	39.20
	ispc	11.60	58.70	43.60	49.60	36.60
	spc	21.80	7.00	21.00	13.00	9.00
Long period	snc	31.20	4.40	11.60	6.70	15.50
	isnc	31.90	45.00	27.60	36.30	51.80
	ispc	23.50	45.90	31.20	43.20	26.30
	spc	13.30	4.70	29.40	13.60	6.30

Note: snc: significant negative correlation; isnc: insignificant negative correlation; ispc: insignificant positive correlation; spc: significant positive correlation.

) revealed that the NDVI exhibited a positive association with TNn, TX90P, and TXx in most regions. Specifically, the areas exhibiting significant positive correlations accounted for 65.7%, 64.6%, and 62.6%, respectively. The significant positive correlation between the NDVI and TNn was predominantly observed in southwestern Qinghai and central Inner Mongolia. Meanwhile, the significant positive responses of the NDVI to TX90P and TXx were mainly distributed in northern Xinjiang, Qinghai, southern Shaanxi, and the area north of the Greater Khingan Mountains. Furthermore, the NDVI showed an inverse relationship with TN10P and DTR across most areas, with negative correlation constituting 66.4% and 54.5%, respectively. Regions where the NDVI exhibited a significantly negative correlation with TN10P were widely distributed in NC, whereas those showing a significantly negative correlation with DTR were more scattered across NC.

In long-term trends, there was an evident spatial variation in the correlation between extreme temperature indices and the NDVI throughout NC (Figure 13 and Table 3). Compared to the three-year timescale, TNn, TX90P, and TXx generally maintained a positive correlation with the NDVI in most regions; however, the spatial extent of this positive correlation had decreased. Specifically, regions where the NDVI and TNn show a positive relationship, TX90P, and TXx accounted for 50.6%, 60.6%, and 58.8%, respectively. In central Inner Mongolia, the significant positive correlation between TNn and the NDVI exhibited a decreasing trend in spatial coverage. Meanwhile, the NDVI was still negatively correlated with TN10P and DTR, mainly in Qinghai, Xinjiang, Northeast China, Shanxi, and Shandong. Among these, the regions demonstrating a significant negative correlation with DTR have expanded.

As for the extreme precipitation index on a 3-year timescale (Figure 14 and

Table 4. Statistics of correlation coefficients between extreme precipitation indicators and the NDVI on different timescales (%).

Timescale	Correlation	CDDs	CWDs	RX1D	RX5D
3a period	snc	5.20	6.40	1.50	2.80
	isnc	43.00	52.10	44.70	47.60
	ispc	49.20	39.50	47.40	44.20
	spc	2.60	1.90	6.30	5.40
Long period	snc	7.30	4.20	0.60	2.40
	isnc	51.20	47.20	29.50	36.00
	ispc	39.60	43.20	53.90	46.00
	spc	1.90	5.40	15.90	15.60

), CDDs and the NDVI demonstrated a positive correlation across most regions, although these areas were relatively scattered. In contrast, CWDs and the NDVI were negatively correlated in most areas, with this pattern primarily observed in Xinjiang and northeastern China. The spatial correlation distributions of RX1D, RX5D, and the NDVI were similar, with the extent of positive correlation regions exceeding that of negative correlations. Notably, significantly positive correlation was predominantly concentrated in southwestern Qinghai, Shanxi, and Shaanxi. In long-time trends (Figure 15), compared to the 3-year period, a negative relationship between the NDVI and CWDs remained prevalent in most regions (Table 4); however, the proportion of areas exhibiting negative correlations has decreased. Furthermore, there has been an increase in positive correlation regions between RX1D, RX5D, and the NDVI, with notably positive associations predominantly found in Qinghai, Inner Mongolia, and the Loess Plateau.

3.5. Vegetation Dynamics in Future Prediction Regression Model

To identify the combinations of key extreme climate factors that exert stable and sustained influences on vegetation dynamics, in this study, we performed a full subset regression analysis based on long-term trends, incorporating nine extreme climate indices and the NDVI (

Table 5. Regression analysis of full subsets between extreme climate indicators and the NDVI in NC.

Variable	Coefficient	Adj-R2	BIC
RX1D	0.37 × 10−1	0.71	−54
DTR	−0.26 × 10−2
TN10P	−0.57 × 10−2
TNn	0.19 × 10−2
TX90P	0.22 × 10−3

). The model outcomes are based on the adjusted R² value. Performing this regression generates models with varying subset dimensions. The highest adjusted R² was obtained when five extreme climate indices, including CDDs, DTR, TN10P, TNn, and TX90P, were used as independent variables. This finding suggests that a linear combination of these five indicators yielded the best model performance, and any addition or removal of other extreme climate indices would diminish the model’s simulation power. Hence, we utilized these five indicators to project future trends in NDVI changes.

3.6. Future Climate Changes and Vegetation Dynamics in NC Based on CMIP6

3.6.1. Model Evaluation

This study employed Taylor diagrams to evaluate the regional performance of CMIP6 global circulation models (Figure 16). Historical data from 1982 to 2018 were compared with model simulation data to comprehensively assess the accuracy and reliability of each model’s performance. The results indicate that the correlation coefficients between model-simulated values and observed values for various extreme climate indices ranged from 0.3 to 0.8, while the RMSE remained below 6.5. The multi-model ensemble mean (MME) demonstrated superior performance compared to individual model simulations. This meant that MME enhanced robustness by mitigating the random errors inherent in individual models. Specifically, its correlation coefficient surpassed that of most single models, and its standard deviation aligned more closely with observed values. Consequently, we utilized the MME to analyze the spatiotemporal variations in extreme climates during the period from 2025 to 2060.

3.6.2. Projected Changes in Future Extreme Climate and Vegetation

This research examines the spatiotemporal trends of extreme climate events from 2025 to 2060 and compares the characteristics of such events between the historical period and future projections (Figure 17). Under both SSP245 and SSP585, the DTR exhibits a continuous decreasing trend compared to the historical period. Additionally, TN10p demonstrates a declining trend in both scenarios, while TNn shows a significant warming tendency. Although the increasing rates of TX90p and RX1d were still positive, they showed a gradual deceleration over time.

Figure 18 illustrates the spatiotemporal variation trends of each extreme climate indicator from 2025 to 2060 under SSP245 and SSP585. It is evident that all indicators exhibit significant spatial heterogeneity. Notably, the variation trends of indicators DTR and RX1D are relatively consistent in both scenarios. For DTR, regions exhibiting decreasing trends are predominantly concentrated in the northwestern area, while increasing trends are mainly concentrated in Shandong and Shanxi Provinces. For RX1D, most regions demonstrate an upward trend, with the highest increase rates being 5.17 mm/a and 5.41 mm/year, respectively. TN10P shows a downward trend in most areas, with significant decreases primarily observed in higher-latitude regions such as Xinjiang and Inner Mongolia. For TNn, both scenarios indicate an upward trend across most of the NC. Under SSP585, the degree to which the upward trend is larger, with notable increases in northern Heilongjiang. For TX90P, most regions exhibit an increasing trend under both scenarios, with maximum increase rates of 0.05 °C/a and 0.07 °C/a, respectively. The increasing trend is particularly pronounced in central Inner Mongolia.

Given that the CMIP6 scenarios simulate the long-term evolution trends of the climate system, this study, based on long-term trends, uses the random forest model to predict future vegetation dynamics under extreme climate conditions. Subsequently, a random forest model was employed to predict future vegetation dynamics under extreme climate conditions. To assess the model’s predictive accuracy, several performance metrics were applied, including R², MAE, MSE, RMSE, and EVS. The results from the training set indicated high predictive accuracy, with R², MAE, MSE, RMSE, and EVS values of 0.9780, 0.0280, 0.0015, 0.0383, and 0.9780, respectively. These results suggested minimal deviation between predicted and actual values, reflecting strong model fitting capability. Furthermore, the model exhibited robust performance on the test set, yielding R², MAE, MSE, RMSE, and EVS values of 0.9670, 0.0342, 0.0022, 0.0469, and 0.9669, respectively. These findings confirm that the model can accurately estimate NDVI values and possesses excellent generalization ability, making it well-suited for predicting new, unseen data (

Table 6. Evaluation metrics of the random forest model.

Statistical Parameter	Random Forest
	Train	Test
R2	0.9780	0.9670
MAE	0.028	0.0342
MSE	0.0015	0.0022
RMSE	0.0383	0.0469
EVS	0.9780	0.9669

).

Under both scenarios, the NDVI exhibited an increasing tendency, with growth rates of 0.007/a and 0.003/a, respectively (Figure 19). However, as radiative forcing intensifies, the growth rate decelerates and the average NDVI value declines. Spatially, under SSP245, the NDVI demonstrates an increasing trend across most regions of NC (Figure 20). Specifically, areas with non-significant increases account for 57.9%, while those with significant increases constitute 3.0%. Areas with non-significant decreases make up 23.7% and areas with significant decreases comprise 15.4%. The decreasing regions are predominantly located in arid zones in the northwest and northeast. Under SSP585, compared to SSP245, the extent of NDVI degradation in NC expands, covering 51.2% of the NC, including 46.9% with non-significant decreases and 4.3% with significant degradation. Notably, degradation trends emerged in the North China Plain, as well as Gansu, Shaanxi, and Shandong under SSP585.

4. Discussion

4.1. Characteristics of Vegetation Dynamics in NC

The results of this study reveal a consistently fluctuating upward trend in vegetation dynamics during the period of 1982–2018 in NC, aligning with previous research findings [40]. Significantly, notable regional disparities were observed in the vegetation dynamics within the study area. Southern Xinjiang, as well as central and western Inner Mongolia, displayed a discernible trend towards degradation. This phenomenon can be attributed to the fragile ecological environment in these regions, where adverse climatic conditions may potentially exacerbate vegetation deterioration [41]. Conversely, regions exhibiting increasing interannual NDVI trends were primarily located in northwestern Xinjiang, eastern Qinghai, the Loess Plateau, and northeastern China. Previous studies have consistently demonstrated that these areas are undergoing concurrent trend of warming and wetting trends [42,43,44,45], a finding corroborated by the results presented in this paper. Furthermore, in recent years, China has initiated numerous ecological restoration initiatives aimed at recovering vegetation, such as the Grain for Green Program, the Natural Forest Protection Project, and the Three-North Shelterbelt Project [46]. These programs have substantially improved the ecological environment and promoted vegetation growth in the regions mentioned. The EEMD results of the NDVI in NC revealed the presence of both 3-year and 6-year cycles, along with a long-term growth trend (Figure 5). Notably, the NDVI showed significant changes in both short-term and long-term trends, consistent with previous studies on vegetation dynamics over multiple timescales in the six northwestern provinces of China [47].

4.2. Characteristics of Extreme Climate Indices Variations in NC

In recent years, there has been a noticeable increase in the amount of attention paid to extreme weather phenomena due to their significant impacts on the economy, society, and ecosystems. Analyses of nine extreme climate indices from 1982 to 2018 revealed a notable increase in both extreme temperatures and precipitation levels in NC, suggesting a trend toward higher humidity. However, spatial variations reveal certain heterogeneity, consistent with previous findings [48]. Although various factors such as topography, altitude, heat islands, volcanic activity, and solar activity influence extreme climate change, large-scale atmospheric circulation is the main driver in mid-latitude regions [27]. This study found that TN10P decreased while TNn increased across most regions, possibly due to strong summer and autumn anticyclones dominated by dry, hot air masses. Additionally, intensified northwest winds hinder the northward movement of moist oceanic air, leading to rising extreme temperatures in NC. In contrast, during winter and spring, stronger cyclones and weakened winter monsoons limit cold air intrusion [49]. The strengthening of the Arctic Oscillation and Southern Oscillation since 1980 has also contributed to this warming trend [50]. A decreasing trend in DTR was observed in northwest China, likely linked to urban heat island effects, increased aerosols concentrations, or land-use changes [51,52]. Extreme precipitation events have significantly increased in regions such as Qinghai, Shaanxi, Shanxi, Inner Mongolia, and Xinjiang. These changes are closely related to the East Asian climate system, together with variations in the Southern Oscillation, Arctic Oscillation, Indian summer monsoon, and North Atlantic Oscillation [53]. In addition, rising temperatures accelerate glacier melting and enhance ocean surface evaporation, thereby increasing high-altitude water vapor concentrations. This elevated moisture level can trigger rainfall and lead to more frequent and intense precipitation events [54]. The decline in precipitation indices in NC may be due to negative winter height anomalies in Siberia and positive summer anomalies in the western Pacific. These anomalies push the West Pacific subtropical high-pressure system southward, dispersing moisture away from NC and reducing precipitation [55].

The evaluation based on the Taylor diagram shows that the MME has a good simulation ability for most climate indicators. However, the correlation coefficient between the MME and the observation is relatively low for the TNn indicator, indicating that there is a certain deviation in the simulation of the interannual fluctuation trend. Nevertheless, it is worth noting that the standard deviation and RMSE of the MME are relatively small, suggesting that the model still has a certain degree of rationality in simulating the variability and overall magnitude of TNn. Hence, its simulation of the long-term average state is relatively reliable. Especially under the high-emission scenario (SSP585), the long-term trend and extreme change direction simulated by it still have reference significance. The spatiotemporal variation trends of the five primary extreme climate indices influencing the NDVI in NC during the period 2025–2060 revealed that all indicators were affected by the exacerbation of global warming. The persistent decline in DTR was primarily driven by rising nighttime temperatures and was comprehensively influenced by the greenhouse effect, alterations in cloud cover, and reductions in aerosols concentrations [56]. The increase in TNn reflected the attenuation of extreme low temperatures due to global warming, which was particularly evident in high-latitude regions such as Xinjiang and Heilongjiang (Figure 18). The rise in RX1D might be attributed to global warming and the monsoon circulation system, especially in the eastern region of the research area where RX1D exhibited a significant increase, potentially modulated by the East Asian summer monsoon [56].

4.3. The Impact of Historical and Future Extreme Climate Changes on Vegetation Dynamics

The spatial relationship results (Figure 12, Figure 13, Figure 14 and Figure 15) indicate that, on a three-year timescale, vegetation growth exhibits stronger positive correlations with TNn, TXx, and TX90P compared to the long-term trend, with a broader spatial extent of positively correlated regions. Except for the southeastern Xinjiang and central to western Inner Mongolia, short-term extremely high-temperature indicators, specifically TX90P and TXx, exhibit a positive correlation with vegetation growth. Southeastern Xinjiang and central-western Inner Mongolia are located in arid areas with desert vegetation. Global warming is projected to exacerbate drought conditions in these arid areas, as soaring temperatures significantly accelerate the evaporation of soil moisture, thereby impairing essential physiological processes in native plant species [57]. Therefore, in arid regions such as southern Xinjiang, ecological restoration strategies should incorporate enhanced irrigation efficiency and water conservation policies. In the humid or semi-humid regions of southern Shaanxi, southern Qinghai, and the Greater Khingan Range, the occurrence of heavy rainfall and consecutive wet days has increased (Figure 11). Elevated temperatures have accelerated the evaporation of residual surface water, fostering environmental conditions favorable to vegetation growth and partially mitigating plant transpiration [36]. Long-term trend analysis further indicates that in regions such as northern Xinjiang, central and western Inner Mongolia, and the Loess Plateau, RX1D and RX5D demonstrate extensive positive correlations with vegetation growth. This can be attributed to the fact that extreme precipitation can stimulate vegetation growth through the continuous enhancement of soil moisture levels in arid and semi-arid areas [58]. Meantime, Jiang, et al. [59]) demonstrated a statistically association between the frequency of extreme precipitation and enhanced vegetation vitality in their study conducted in NC. Their findings indicate that both short-term high temperatures and long-term precipitation are pivotal in promoting vegetation recovery across various regions within NC.

Over the next 36 years, the impact of extreme climates on vegetation shows clear regional differentiation and scenario dependence (Figure 19 and Figure 20). The more extensive increase in the NDVI under SSP245 highlights the positive effect of mild climate change on vegetation. In contrast, under SSP585, the synergistic effects of multiple extreme climate events lead to a decrease in the NDVI in the central and western part of the NC. Vegetation degradation in northern Xinjiang is primarily driven by the synergistic effects of a decrease in DTR and TN10P (Figure 18). The reduction in DTR leads to an increase in plant respiration consumption and a decrease in the accumulation of net photosynthetic products. Meanwhile, the decline of TN10P accelerates the melting of winter snow and weakens the water supply in spring. The two factors work together to intensify drought stress, causing a decline in the resilience of ecosystems that rely on snowmelt water. In the central region of the study area (e.g., Gansu Province), the observed NDVI decline results from the combined influence of reduced DTR, increased TX90P, and a slight decrease in RX1D. The narrowing DTR continues to impair photosynthetic efficiency, while elevated TX90P intensifies evaporative demand. Additionally, the marginal decrease in RX1D reduces precipitation effectiveness. The cumulative impact of these three factors gives rise to a “high temperature–low water” stress regime, which may hasten the onset of vegetation degradation, particularly in arid zones such as the Hexi Corridor. Hence, these results suggest that high-emission pathways may pose greater threats to ecosystem stability [60]. Under SSP245 and SSP585, the notable rise in TNn in Heilongjiang will alleviate low-temperature stress and promote the prolongation of the vegetation period in high-latitude regions. However, the simultaneous increase in TX90P and RX1D may offset some of the benefits. The wetland ecosystem in Heilongjiang is at risk of waterlogging due to intensified extreme precipitation [61]. These findings emphasize the necessity of future ecological management strategies to account for regional climate sensitivity. Future research should aim to integrate climate–vegetation models with socioeconomic data to optimize spatial prioritization for large-scale ecological restoration initiatives, including China’s “Three-North Shelter Forest Program”.

4.4. Limitations and Future Perspectives

In this study, the meteorological data obtained from northern stations were spatially interpolated to generate continuous surface data. However, due to the limited coverage of the station network, the interpolation method has inherent limitations, resulting in discrepancies between the interpolated results and actual ground-level meteorological conditions. Hence, future efforts should focus on improving the accuracy of datasets for extreme climate analysis through denser observation stations. Additionally, this study considered only the impact of extreme weather on the NDVI; however, the NDVI is influenced by a range of factors, including natural aspects such as topography and soil characteristics, as well as anthropogenic factors like GDP and population density. Henceforth, future research should aim to comprehensively investigate these various influences. Moreover, although the current analysis focused on regional-scale vegetation trends, different land cover types may respond uniquely to extreme climate events. These variations stem from differences in physiological traits, management practices, and ecosystem resilience. Therefore, future research should systematically compare vegetation type-specific sensitivities to support targeted adaptation strategies. Finally, future projections remain uncertain. While CMIP6 has improved in terms of model resolution and physical processes, uncertainties among multi-model ensembles still exist. Regional responses to climate change also vary significantly. For example, projected precipitation trends in arid regions show large discrepancies, increasing regional uncertainty. Although model improvements have enhanced projection accuracy to some extent, uncertainties from model differences and regional variability still require further research and quantification.

5. Conclusions

This research employed remote sensing and meteorological data to analyze the tendency of variation in the NDVI in NC from 1982 to 2018, as well as the influence of extreme climate indices on the NDVI. Additionally, it projected the spatiotemporal trends of future extreme climate conditions and the NDVI under SSP245 and SSP585. The primary conclusions are summarized as follows.

From 1982 to 2018, the NDVI in NC generally exhibited an upward trend at a rate of 0.003 per year; however, a decrease was observed in central and western Inner Mongolia, the Hexi Corridor, and southern Xinjiang, indicating the persistence of localized ecological degradation. At both short-term (3-year) and long-term temporal scales, the NDVI in most regions showed a negative response to extreme temperature indicators like DTR and TN10P, but a positive correlation with TNn, TX90P, and TXx. For extreme precipitation indices, the NDVI responded positively to RX1D and RX5D over both short- and long-term periods in most areas. Additionally, the NDVI demonstrated a positive association with short-term CDDs but a negative correlation with long-term CDD. The effects of CWD and CDDs on the NDVI were found to be contrasting. Overall, the spatial extent to which vegetation growth was positively influenced by short-term high temperatures was relatively large, whereas the area where long-term extreme precipitation promoted vegetation growth was more extensive.

Based on long-term trend analysis, full subset regression reveals that DTR, TN10P, TNn, TX90P, and RX1D exert the most significant influence on vegetation in NC. The projected temporal variation trends of these five indices from 2025 to 2060 align with historical changes. In high-latitude regions of NC, both low and high temperatures rise, while precipitation increases in the central and eastern regions. Under both SSP245 and SSP585, the future growth rate of the NDVI is expected to decline markedly compared to the historical period, with an increasing number of regions experiencing negative changes as radiation intensity rises.