Remote Sensing Identification and Stability Change of Alpine Grasslands in Guoluo Tibetan Autonomous Prefecture, China

Abstract

Alpine grasslands, a crucial component of the Qinghai–Tibet Plateau, play a vital role in maintaining ecological barriers and facilitating sustainable development, and the exact stability change is also the key to coping with climate change and implementing ecological protection projects. The purpose of this study was to identify the spatial and temporal distribution of multi-stage alpine grassland and explore its inter-annual distribution and growth stability. The Guoluo Tibetan Autonomous Prefecture, China (hereinafter referred to as Guoluo), where alpine grassland is widely distributed, was selected as the research area. Long-term stable grassland samples constructed using the Mann–Kendall–Sneyers mutation test method were analyzed alongside random forest classification to identify multi-stage grassland distribution trends from 1990 to 2020. Based on the Fractional Vegetation Cover (FVC) and coefficient of variation (C_v), spatial and temporal changes in grassland quality and their driving factors were discussed. The results show the following: (1) Remote sensing grassland extraction, based on the establishment of long-term stable grassland samples and random forest classification, demonstrated high accuracy and reliability, with OA and Kappa coefficients consistently above 0.89 and 0.77, and PA and UA maintained consistently at approximately 0.9. (2) The distribution of grassland in Guoluo corresponded to the spatial patterns determined by the natural geographical environment, showing a gradual trend from high-cover grassland in the southeast to low-cover grassland in the northwest. The proportion of medium and high-cover grasslands slightly increased, indicating an improvement in grassland quality. However, the encroachment and degradation caused by human activities and climate change resulted in a slight decrease in the proportion of grassland area compared with 1990. (3) Despite the overall grassland ecosystem still having relative stability, local grassland quality changes dramatically, mainly in the north of Maduo County. And significant fluctuations in the area of grassland quality were noted over the last two decades, suggesting potential degradation in ecosystem stability. Climate change and human activities were identified as primary drivers of these changes. Climate change is dominant in the alpine region. The low-warming region is dominated by human activities. These findings offer essential insights for the planning and implementation of alpine grassland ecosystem protection and restoration initiatives and also have important value for exploring the evolution law of alpine grassland ecosystems.

1. Introduction

Alpine grasslands encompass various vegetation types, including meadows, grasslands, swamp–aquatic grasslands, cushion grasslands, sparse grasslands, and scrub grasslands, typically found in cold areas at higher elevations [1]. However, their susceptibility to factors like climate change, overgrazing, and land use changes poses significant challenges to their stability [2]. Recovery from disturbances in alpine grassland ecosystems is often slow, and large-scale degradation can further impede the sustainable development of regional ecological environments and social economies.

In China, alpine grasslands predominantly occur in the central and southern Qinghai–Tibetan Plateau, Pamir Plateau, and Tianshan, Kunlun, and Qilian Mountains. These areas boast abundant water and grass resources, making them crucial natural pastures in China. Moreover, China’s alpine grassland regions serve as the headwaters of important rivers in East Asia and South Asia. Changes in alpine grasslands can profoundly affect downstream water resources and ecological security [3].

Given these considerations, monitoring changes in alpine grasslands has been a primary focus of scientific research. Many scholars have investigated change identification [4,5], attribution [6], and potential impacts [7,8,9] on alpine grasslands [10,11], employing both field investigations and remote sensing techniques. For instance, Zhang et al. [4] used high-resolution images to identify various types of degraded grasslands in the source region of the Yellow River and summarized the quantity and distribution of these degraded grassland types. Similarly, Chen et al. [7] analyzed the correlation between vegetation change, climate, and human activities in the three-river headwaters region using NDVI data from 1995 to 2014. Such studies have provided valuable insights into the protection and management of alpine grasslands. However, most of the previous research has focused on identification and change analysis in regional grasslands, with detailed spatial change detection being relatively scarce.

From an ecosystem perspective, the stability of elements is crucial for ensuring ecosystem resistance to external changes and damage [12]. However, most previous studies on alpine grassland stability have primarily focused on soil microorganisms and vegetation diversity [13]. There has been a lack of unified detection of alpine grassland stability on a large scale.

Most prior studies investigating the mechanisms influencing alpine grassland stability [14,15] have centered on the role of individual factors, neglecting a systematic discussion of the various factors affecting stability. Previous research [16] has provided limited insights into the spatial distribution, stability, and driving factors of stability change in alpine grasslands. These factors are crucial for understanding the evolution of alpine grassland ecosystems and implementing effective ecological protection measures.

In this study, multi-stage alpine grassland was extracted using remote sensing technology. Detailed detection and attribution analysis of alpine grassland ecosystem stability were conducted. The Guoluo Tibetan Autonomous Prefecture, hereinafter referred to as Guoluo, characterized by alpine grassland as the predominant land cover, was selected as the research area. Remote sensing was employed to identify the distribution range of alpine grasslands from 1990 to 2020, and the variation coefficient of multi-stage alpine grassland coverage was calculated. The relationship between the spatial distribution of alpine grasslands and changes in grassland cover, as well as the spatial distribution of grassland stability and related factors, was examined to provide insights for evolutionary studies of alpine grasslands and the formulation of ecological protection strategies.

2. Materials and Methods

2.1. Study Area

Guoluo, situated in the eastern Qinghai–Tibet Plateau (Figure 1a), serves as the core area of the Yellow River [17]. The terrain slopes from northwest to southeast, with the Yellow River system at its center, gradually rising on both sides and boasting an average altitude exceeding 4200 m. Characterized by a typical plateau continental climate, the region experiences an average annual temperature of −4 °C and lacks an absolute frost-free period throughout the year. Alpine grasslands dominate the landscape, covering 88% of the land area (Figure 1b). The available grassland area spans approximately 6.26 million ha, accounting for 92.6% of the natural grassland area, and forms a vital part of the principal pastoral areas on the Qinghai–Tibet Plateau [18]. Changes in alpine grasslands are intricately linked to the ecological barrier, the status of the livestock industry on the Qinghai–Tibet Plateau, and the high-quality development of the natural, economic, and social aspects of the Yellow River Basin.

2.2. Data and Preprocessing

The data utilized in this study primarily consisted of remote sensing, land cover, terrain, meteorological, and basic geographic data. To ensure temporal continuity, consistency, and spatial resolution accessibility, multispectral image data collected by Landsat 5 and Landsat 8 satellites were chosen for long-time series remote sensing analysis, particularly for alpine grassland recognition. During data acquisition, consideration was given to the phenological characteristics of the study area and the impact of cloud cover on image quality. Data products were obtained through the Google Earth Engine (GEE) platform after undergoing radiometric calibration, atmospheric correction, and ortho-correction from June to August of each year across seven stages: 1987–1991, 1993–1997, 1999–2001, 2004–2006, 2009–2011, 2014–2016, and 2019–2021. Utilizing the Cloud Quality Detection Band (QA) and Fusion Mosaic Function of GEE, seven valid datasets representing the years 1990, 1995, 2000, 2005, 2010, 2015, and 2020 were synthesized. Furthermore, high-resolution Google Earth satellite images from 2020 and the surrounding years were synchronously accessed for human–computer interaction, facilitating the refinement of classification samples based on existing data products.

The land cover data utilized were sourced from the ESA WorldCover2020 dataset [19], released by the European Space Agency, primarily for generating initial samples for alpine grassland recognition. A total of 1278 initial training samples and 907 verification samples were randomly selected within the research area based on the ESA WorldCover2020 data (

Table 1. Sample information.

Classifications	Number of Training Samples				Number of Validation Samples
	Initial	Visual Revisions	Filters	Sample Rejection Rate	Initial	Visual Revisions	Filters	Sample Rejection Rate
Grassland	967	40	927	0.041	596	29	567	0.049
Forest	100	6	94	0.060	203	13	190	0.064
Water	109	16	93	0.147	55	7	48	0.127
Others	102	7	95	0.069	53	2	51	0.038
Total	1278	69	1209	0.054	907	51	856	0.056

).

Topography and geomorphology play crucial roles in determining the distribution of grassland resources [20]. Therefore, in remote sensing-based grassland extraction, features such as elevation, slope, and slope direction were included in the feature set. Elevation data were obtained from the Digital Elevation Model (DEM) in the NASA SRTM Digital Elevation 30 m dataset [21], while slope and slope direction were calculated using DEM-based terrain data in the GEE.

Meteorological data primarily consisted of temperature and precipitation, derived from the monthly precipitation dataset of China with a 1 km resolution (1901–2021) and the monthly average temperature dataset of China with a 1 km resolution (1901–2021), developed by Peng et al. [22,23]. Data from 1988 to 2021, extracted based on the boundary of the study area, were utilized as factors influencing grassland stability.

Geographic data, including administrative division boundaries, residential site distributions, highway distributions, water system distributions, population density, and soil types of Guoluo, were obtained from the National Geographic Information Resources Directory Service System 1:1,000,000 national basic geographic database [24]. Additionally, the Lake Dynamic Dataset of the Qinghai–Tibetan Plateau (V1.0) (1984–2016) [25] was employed to delineate the study area’s scope and investigate the relationship between alpine grassland stability and the proximity of highways, residential sites, water sources, population density, and soil types.

2.3. Method and Process

2.3.1. Random Forest Classification

Random forest classification is a supervised integrated machine learning algorithm widely recognized for classifying remote sensing and geographic data [26,27]. It offers considerable adaptive advantages in classifying land cover types, clouds, and fog [28,29]. Therefore, this study directly selected the random forest classification method by utilizing the built-in GEE function Classifier.SmileRandomForest. A total of 263 decision tree training classifiers were selected by referencing [30] to identify alpine grasslands, with the other parameters set to the default values of GEE functions. The extraction and construction processes for long time-series classification samples and feature sets as inputs and tests for model training are as follows:

(1) Construction of stable classification samples for long time-series.

Based on the ESA World Cover 2020 data [19], initial sample points representing land cover types in 2020 were randomly generated. High-resolution satellite imagery from Google Earth was utilized to refine and create a benchmark sample through human–computer interaction (Table 1).

Due to the widespread absence of historical high-resolution remote sensing image data, visual interpretation based on medium- and low-resolution images requires the knowledge and experience of participants. However, controlling the quality of time series samples becomes challenging when obtained solely through human–computer interaction. Therefore, leveraging a benchmark sample from 2020, this study introduced the Mann–Kendall Sneyers mutation test (MKS) [31] to construct stable long-term series classification samples.

MKS is a non-parametric statistical method that eliminates the need to assume the probability distribution of samples. It directly assesses the monotonicity of time series samples by calculating rank differences and symbols to facilitate trend analysis and detect mutation points within time series data [32]. The MKS algorithm, characterized by its simplicity and efficiency, exhibits high adaptability to missing values and nonuniform time intervals [33]. Widely employed in analyzing change trends and mutations in meteorological and hydrological fields, MKS effectively mitigates the impact of outliers in time series data and enhances change detection accuracy [33]. The specific calculation process for MKS is as follows:

For a time series X $\left\{x_1,x_2,\dots \dots \dots ,x_n\right\}$ with n sample sizes, construct an order column:

$$S_k=\sum _1^k{R}_i,2\le k\le n$$

(1)

where S_k represents the cumulative number of x_i greater than x_j ($1\le i\le j$).

$$R_i=\left\{\begin{array}{c}1,x_i>x_j\\ 0,x_i\le x_j\end{array}\right.\left(j=1,2,\dots ,i\right)$$

(2)

Under the assumption of random independence of time series, the statistics are defined as:

$$U{F}_k=\frac{S_k-E_{\left(S_k\right)}}{\sqrt{Var\left(S_k\right)}}\hspace{1em}k=1,2,3,\dots \dots ,n$$

(3)

where, if k = 1, $U{F}_1$ = 0; $E_{\left(S_k\right)}\mathrm{and}Var\left(S_k\right)$ are the mean and variance of $S_k$, respectively. ${\{x}_1,x_2,\dots \dots ,x_n\}$ are independent of each other and have the same continuous distribution, and $E_{\left(S_k\right)}$, $Var\left(S_k\right)$ can be obtained from Equations (4) and (5):

$$E_{\left(S_k\right)}=\frac{n\left(n-1\right)}{4}$$

(4)

$$Var\left(S_k\right)=\frac{n\left(n-1\right)\left(2n+5\right)}{72}$$

(5)

where n represents the number of samples of time series X.

In theory, $U{F}_k$ is a standard normal distribution, which is determined by the time series ${\{x}_1,x_2,\dots \dots ,x_n$} sequence of the calculated statistics. If $U{F}_k$ exceeds the confidence interval, it indicates that the time series has a prominent change trend. In this study, the significance level α = 0.05 corresponds to a confidence interval of 1.96.

Then, the time series X is constructed in reverse order $x_n,x_{n-1},\dots 1$, and the above process is repeated to obtain $U{B}_k$, such that:

$$U{B}_k=-U{F^{\prime }}_k\hspace{1em}k=n,n-1,\dots ,1$$

(6)

If $U{B}_k$ and $U{F}_k$ intersect in the confidence interval, the time series has a mutation. The moment corresponding to this intersection marks the onset of the mutation. The direction of the mutation is indicated by the positive or negative value of $U{F}_k$.

The Normalized Difference Vegetation Index (NDVI) exhibits strong differentiation among various ground feature types [34]. MKS is capable of effectively detecting sudden changes in time series NDVI resulting from disasters, climate change, and human activities. These abrupt changes often correspond to alterations in ground types. Therefore, NDVI was selected as an index for MKS to construct stable classification samples over a long time series. NDVI values from multi-stage remote sensing data were computed separately (Equation (7)). Subsequently, time series NDVI samples were extracted based on the benchmark sample from 2020. MKS was then applied to eliminate sample points with mutations. The final sample selection results are presented in Table 1 and Figure 2.

$$NDVI=\frac{\left(NIR-Red\right)}{\left(NIR+Red\right)}$$

(7)

where NIR represents the near-infrared band and red represents the infrared band.

(2) Construction of classification feature set.

The feature set extracted from alpine grasslands consisted of spectral reflectance, spectral indices, spectral transformations, and terrain features [26]. Spectral reflectance captures the spectral response of various ground objects and exhibits strong differentiation between different vegetation types and bare land [35]. This study utilized six bands: blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2. By mathematically combining multiple bands, spectral index features highlight significant spectral characteristics of ground objects, thereby enhancing classification effectiveness and accuracy [36]. The four spectral characteristic indices—normalized vegetation index (NDVI), ratio vegetation index (RVI), enhanced vegetation index (EVI), and normalized differential water index (NDWI)—were directly selected based on the existing literature [26] and are detailed in

Table 2. Information on taxonomic characteristic elements of alpine grassland.

Feature Name	Feature Index	Algorithm	Note
Index Features	NDVI	(NIR − R)/(NIR + R)	NDVI: normalized differential vegetation index; NDWI: Normal Differential Water Index; EVI: Enhanced Vegetation Index; RVI: ratio vegetation index. B, G, R, and NIR correspond to the blue, green, red, and near-infrared bands in the multispectral image. Y represents the output value of the KT transformation, X is the input value, c is the transformation matrix, and a is a constant added to avoid negative numbers.
	NDWI	(G − NIR)/(G + NIR)
	EVI	2.5/(NIR − R)/(NIR + 6 × R − 7.5 × B + 1)
	RVI	NIR/R
Tasseled Cap Transformation	brightness	Y = cX + a
	wetness
	greenness

. NDVI is widely employed in vegetation classification, primarily to differentiate between vegetation and non-vegetation land cover types [26]; RVI is tailored to enhance the distinction between grasslands and forestlands [26]; EVI aids in vegetation identification by correcting for soil and atmospheric influences, effectively mitigating their impact on vegetation extraction [26]; NDWI significantly aids in water body identification [37]. The spectral transformation feature primarily involves the linear transformation of the original image using the Tasseled Cap Transformation (TCT) to reduce data volume and extract feature values. Among these, TCT extraction of soil brightness and vegetation greenness introduces a new band reflecting maximum variance information in the original data, significantly distinguishing forested areas (Table 2) [38]. Topographic features enhance classification information and differentiation [39], typically encompassing elevation, slope, and slope direction. The classification included 15 variables across four categories.

2.3.2. Precision Evaluation

The accuracy of remote sensing classification is primarily evaluated using relevant indicators calculated based on a confusion matrix [40,41,42]. This method was also employed in this study, and the specific indicators and algorithms are listed in

Table 3. Calculation method for each indicator for accuracy verification.

Index	Algorithm	Note
Producer’s Accuracy	$PA=TP/(TP+FN)$	N represents the total number of samples. TP stands for true positives, which refers to the number of positive samples that have been correctly identified. FN stands for false negatives, which refers to the number of positive samples that have been missed. FP stands for false positives, which refers to the number of negative samples that have been incorrectly identified as positive. PRE represents the random accuracy.
User’s Accuracy	$UA=TP/(TP+FP)$
Overall Accuracy	$OA=\sum TP/N$
Kappa	$Kappa=(OA-PRE)/(1-PRE)$ $PRE=\frac{\left(TP+FP\right)\cdot \left(TP+FN\right)+\left(TN+FP\right)\cdot \left(TN+FN\right)}{(TP+FP+TN+FN{)}^2}$

.

2.3.3. Grassland Quality Division

To conduct a unified analysis of grassland growth, this study calculated the grassland cover and divided their quality according to the fractional vegetation cover (FVC). Specifically, the method of synthesizing the maximum NDVI value of the growing period was used to represent the NDVI value of the grassland in the current year, and the 5–95% pixel dichotomy was used to calculate the vegetation cover [43] to obtain the vegetation cover data of Guoluo over seven stages with an interval of five years from 1990 to 2020. The vegetation cover was calculated as follows:

$$FVC=\frac{NDVI-{NDVI}_{soil}}{{{NDVI}_{veg}-NDVI}_{soil}}$$

(8)

where ${NDVI}_{soil}$ is the minimum value of a pure soil pixel, which is close to zero in theory, and ${NDVI}_{veg}$ is the maximum value of a pure vegetation pixel, which is close to one in theory. Considering the actual distributions of vegetation and bare land in Guoluo, pixel values of 0.05 and 0.95 were used to replace ${NDVI}_{soil}$ and ${NDVI}_{veg}$ according to the classification results.

According to Reference [44], this study divided 5–20% FVC into low-cover grasslands, 20–50% FVC was divided into medium-cover grasslands, and >50% FVC was divided into high-cover grassland. In addition, <5% of FVC was classified as “other”.

2.3.4. Stability Analysis of Alpine Grassland

Based on the definition of grassland ecosystem stability [45], this study utilized the coefficient of variation (C_v) to assess grassland stability across the seven periods. C_v is a statistical index employed to gauge the magnitude of data changes in a time series, accurately reflecting variations in data over time [46]. The regional stability of grassland ecosystems was evaluated by computing the C_v of grassland cover over the past 30 years. The C_v was calculated as follows:

$$\overline{X}=\frac{1}{n}\sum _{i=1}^n{x}_i$$

(9)

$$C_v=\frac{S}{\overline{X}}\times 100\mathrm{\%}=\frac{1}{\overline{X}}\sqrt{\frac{\sum _{i=1}^n(x_i-\overline{X}{)}^2}{n-1}\times 100\mathrm{\%}}$$

(10)

where $x_i$ represents the cover value and $n$ represents the number of periods. In this study, $n$ is 7. The greater the $C_v$ value, the greater the inter-annual change in grassland cover and the weaker the stability of the grassland.

To explore the correlation between the intensity of grassland change and the relevant driving forces, this study employed a geodetector [47,48,49] to investigate the mechanisms influencing grassland change. A geodetector is a statistical tool utilized for detecting spatial differentiation and its driving influences. Factor detection was utilized to interpret a certain independent variable in the spatial differentiation of the dependent variable. The q value serves as the primary factor detection index in the geodetector and is utilized to measure the impact factor explaining the spatial differentiation of the dependent variable (see Equation (11)). Interaction detection compares the q value of the factor detection of a new layer formed by the superposition of two factors with the value of the single-factor detection result to detect the effect of different combinations of two factors on the dependent variable. This method was used to examine whether the interpretation of the dependent variable is enhanced, weakened, or independent compared to the single factor under the joint action of two factors. To determine whether the combination of one influencing factor and another had a significant influence on the spatial distribution of grassland stability, the F-statistic was employed (see Equations (12) and (13)). If the influence of these two factors on the spatial distribution of grassland stability was significant, it was labeled as “Y”; otherwise, it was marked as “N”.

$$q=1-\frac{\sum _{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}$$

(11)

where L is the classification or subdivision of grassland variation coefficient or influence factor; $N_h$ and ${\sigma }_h^2$ are the number of units and variance of layer h, respectively; and N and ${\sigma }^2$ are the total number and variance of units in the study area, respectively. The q statistic is in the range of 0–1. The closer q is to 1, the stronger the explanation of this factor for the variation in the C_v of grassland.

$$F=\frac{N{x}_1\left(N_{x2}-1\right)SS{W}_{x1}}{N_{x2}\left(N_{x1}-1\right)SS{W}_{x2}}$$

(12)

$$SS{W}_{x1}=\sum _{h=1}^{L1}{N}_h{\sigma }_h^2,\hspace{1em}SS{W}_{x2}=\sum _{h=1}^{L2}{N}_h{\sigma }_h^2$$

(13)

where F is the statistical value; $N{x}_1$ and $N{x}_2$ represent the sample sizes of the two factors x₁ and x₂, respectively; SSW_x1 and $SS{W}_{x2}$ represent the sum of the in-layer variances of the layers formed by x₁ and x₂, respectively; and L1 and L2 represent the number of layers of variables x₁ and x₂, respectively. After establishing the null hypothesis H0: $S{W}_{X1}$ = $SS{W}_{X2}$ and calculating the F value, if F is greater than the $F_{\alpha }$ obtained by looking up the table at the significance level α = 0.05, the null hypothesis H0 is rejected. This indicates that there are significant differences in the influence of the two factors x₁ and x₂ on the spatial distribution of attribute Y. In this present study, these two factors significantly affected the spatial distribution of grassland stability.

According to existing research results [50,51,52] and the actual situation of the research area, among the background factors which determine the development environment of grassland, and the factors affecting the growth of grassland, a total of 10 factors were selected to explore the mechanism of influence on grassland stability change. The background factors include altitude, slope, slope direction, and soil type, and the factors affecting the growth of grassland include temperature, precipitation, distance from residential areas, distance from roads, and distance from water systems. Because geodetector processing requires discrete data, it is necessary to reclassify the continuous factors into 10 categories using the natural breakpoint method during the construction process and assign each category a unique grade value to ensure the applicability of data in the construction process of the data model. To ensure that the sampling points and C_v of the input geodetector were evenly distributed in space, the C_v was divided into seven categories using the natural breakpoint method. A total of 28,000 sampling points were sampled using the stratified averaging method to achieve uniform sampling of each C_v layer. Finally, with the assistance of GIS analysis software, the grade values of the 10 key factors were sampled for each sampling point, and the complete dataset of the 10 key factors of all sampling points was imported into the geographic detector for subsequent analysis

3. Results

3.1. Grass Extraction Accuracy

Table 4. Classification accuracy.

Years	PA				UA				OA	Kappa
	Grassland	Water	Forest	Others	Grassland	Water	Forest	Others
1990	0.996	0.958	0.711	0.569	0.879	0.979	0.985	1.000	0.922	0.835
1995	0.872	0.979	0.977	0.848	0.872	0.979	0.977	0.848	0.893	0.770
2000	0.995	0.958	0.758	0.569	0.891	1.000	0.986	0.935	0.91	0.817
2005	0.993	0.958	0.784	0.627	0.905	0.958	0.980	0.941	0.922	0.837
2010	0.989	0.979	0.779	0.647	0.903	0.979	0.987	0.892	0.905	0.796
2015	0.989	1.000	0.784	0.608	0.902	1.000	0.980	0.912	0.921	0.834
2020	0.986	1.000	0.805	0.608	0.907	0.980	0.981	0.886	0.924	0.841

displays the accuracy of the confusion matrix test for grassland extraction based on random forest every five years from 1990 to 2020. The accuracy of grassland extraction varied over the six phases of image processing, yet the overall accuracy (OA) and Kappa coefficients remained consistently higher than 0.89 and 0.77, respectively. This suggests that the samples tested by MKS can reliably support remote sensing extraction of long-time-series grasslands. Although classification performance varied for different objects, water was the most easily recognizable due to its distinct image features and achieved high accuracy. Grassland classification accuracy, as indicated by the producer’s accuracy (PA) and user’s accuracy (UA), was slightly lower than that of water but remained stable at approximately 0.9. The random forest classification demonstrated high accuracy and reliability in grassland extraction, which makes the generated data suitable for subsequent grassland analysis.

3.2. Grassland Spatiotemporal Distribution and Change

Grasslands in Guoluo exhibited wide distribution, maintaining a consistent area between 66,936 and 67,932 km² over the last 30 years, covering approximately 88% of the total area of Guoluo (refer to Figure 3 and Figure 4). Undistributed grassland areas were primarily situated in high-altitude mountains in the middle-east of Maduo County, west of Maqin County, and south of Dari County. High-cover grasslands were predominantly distributed in the southeast, while those with medium- and low-cover were concentrated in the northwest. Maduo County exhibited the widest distribution of medium- and low-cover grasslands. The overall pattern of grassland cover in Guoluo featured high coverage in the southeast and low coverage in the northwest, consistent with the natural geographical spatial pattern shaped by the hydrothermal environment. However, spatiotemporal variations in category and quality were also observed.

The first change pertains to categories.

Table 5. Transition matrix of land cover types in Golok State from 1990 to 2020/km².

		2020				Total for 1990
		Grassland	Open Water	Forest	Others
1990	Grassland	66,632.9	251.394	223.0178	829.653	67,936.97
	Open water	144.4463	2072.861	6.65925	97.4955	2321.462
	Forest	565.9755	14.48625	430.6035	0.7005	1011.766
	Others	554.49	84.51525	1.71	2872.557	3513.272
	Total for 2020	67,897.82	2423.256	661.9905	3800.406	74,783.47

illustrates the transfer matrix of land cover types in Guoluo from 1990 to 2020. Different land cover types in Guoluo have experienced varying degrees of change over the last 30 years. Among these changes, the areas of grassland and forest land have decreased, while the areas of water and other types of ground objects have increased. Specifically, due to urban expansion and the development of public infrastructure such as highways, some grasslands have been converted into construction land. Additionally, the combined effects of human activities and climate change may have led to the degradation of certain grasslands, transforming them into bare land. Furthermore, rising water levels in plateau lakes have caused the transformation of some areas into water bodies [53].

The second change concerns the quality of grasslands. Based on the relative changes in grassland cover in Guoluo over the last 30 years (Figure 4), it is evident that the proportion of high grassland coverage in Guoluo has been consistently large, exceeding 60% in each period. The proportion of grasslands with medium cover has fluctuated by approximately 20%, while the proportion of low-cover grassland has fluctuated by around 10%. The combined proportion of grasslands with medium and high cover has gradually increased over time, indicating an overall improvement in grassland growth. Notably, the area of low-cover grassland initially increased and then decreased, reaching a minimum value in 2015. Conversely, the medium grassland cover showed a decreasing trend, whereas the area of high-cover grassland initially decreased and then increased, peaking in 2015. Figure 5 and Figure 6 display the spatial distribution of the number of changes in grassland quality and the statistical results of grassland quality changes during different periods in Guoluo. Regions experiencing changes in grassland quality (Figure 5) were mainly concentrated in the western part of Guoluo, with greater intensity observed in the northern part of Maduo County. Another concentration of changes was observed in the eastern region of Guoluo, characterized by a more linear distribution. These observations align with previous studies on the range and susceptibility of grassland degradation in Guoluo [13], validating the reliability of the extraction results.

The analysis of the time series of changes in grassland quality (Figure 6) reveals a fluctuating trend from 1990 to 2020, characterized by an initial increase followed by a decrease. The change was relatively stable from 1990 to 2005, with more pronounced fluctuations observed after 2005. The highest value was recorded from 2005 to 2010, covering an area of 6157 km², which accounted for 9.4% of the total grassland area in Guoluo during the same period. The minimum value occurred from 2015 to 2020, covering an area of 1873 km². This trend indirectly suggests a potential decrease in the stability of grassland ecosystems after 2005.

3.3. Evaluation and Driving Force Analysis of Grassland Stability in Guoluo

The calculation results for grassland C_v in Guoluo during 1990–2020 are shown in Figure 7. Over the last 30 years, the C_v of vegetation cover in most regions of Guoluo was relatively small, with areas exhibiting large C_v mainly concentrated in the north and northwest of Maduo County. Additionally, high-altitude mountainous areas in the middle and west of Dari County also displayed larger C_v values. Furthermore, other regions with elevated C_v values were primarily situated along rivers and residential areas, demonstrating a linear and clumped distribution pattern with strong aggregation (Figure 7a,b). Analyzing the histogram of the C_v distribution (Figure 8), it is observed that C_v values for over 90% of the regions were less than 0.2, with the largest proportion falling between 0 and 0.15. This indicates that the grassland in Guoluo remained relatively stable, with C_v trending toward lower values, representing slight, normal fluctuations. However, in a few areas, significant changes in grasslands may have occurred due to abnormal local changes in environmental factors, warranting further exploration.

Table 6. Factor probe results.

Factors	Precipitation (X1)	Temperature (X2)	Residential Area Distance (X3)	Road Distance (X4)	Population (X5)	Water Distance (X6)	Soil (X7)	Altitude (X8)	Slope (X9)	Aspect (X10)
q value	0.272	0.149	0.113	0.059	0.212	0.003	0.237	0.096	0.102	0.006
p value	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000

displays the outcomes of the single-factor detection analyzing the driving forces behind grassland change in Guoluo using the geodetector. The order of significance for each factor influencing the variation in grassland C_v, from most to least impactful, is as follows: precipitation > soil type > population density > temperature > distance from settlements > slope > elevation > distance from roads > slope direction > water distance. Notably, precipitation, soil type, and population density exhibited the most substantial explanatory power for grassland variation, with q values all surpassing 0.2. Next, the explanatory power decreased by approximately 0.1 for temperature, residential distance, slope, and elevation. Conversely, road distance, slope direction, and water distance exerted minimal influence on grassland variability, with q values below 0.1. All factors in the table demonstrated p values lower than 0.01, confirming the statistical significance of the results.

Figure 9 illustrates the impact of these factors on grassland variability through interactive detection. The interaction between two factors yielded more pronounced differences compared to single-factor effects. Moreover, the coupling of a high-impact single factor with other factors further amplified its explanatory power for grassland variation. For instance, the interaction between soil (X7) and altitude (X8) yielded a q value of 0.326, while precipitation (X1) combined with all other factors resulted in a q value of around 0.3. This underscores the enhanced explanatory capacity of two-factor interactions in elucidating grassland stability changes. Additionally, the F-test outcomes with a significance level of 0.05 (Figure 10) further validate the coupling of population density (X5) with temperature (X2), residential distance (X3) with road distance (X4), soil type (X7) with temperature (X2), settlement distance (X3), road distance (X4), population density (X5), and water distance (X6), all passing the F-test at the 0.05 significance level.

4. Discussion

Using Guoluo as the study area, the spatial distribution of grasslands was extracted and their stability was evaluated using Landsat remote sensing images from the growing season over the past 30 years. In this study, the concept of grasslands was relatively narrow, excluding grasslands growing in mixed areas of grass and forest, and grass growing on the surface of shallow water. Therefore, the final grassland distribution results may differ from those of grassland distribution in a broader sense.

In the classification process, based on existing land use/land cover data products [19], high-resolution remote sensing data, visual interpretation, and MKS [31], a stable time series sample was constructed to address the problem of obtaining an early sample. Widely used classification methods [26,27] and classification features [26,35,36,37,38,39] were selected. Although the uncertainty of remote sensing may cause fluctuations in classification accuracy over different periods, the accuracy remained relatively stable overall, providing reliable background data for analyzing grassland stability. With the advancement and application of remote sensing science and technology, the optimal methods and construction of features for alpine grassland extraction need further examination to refine existing results.

Based on the grassland distribution data, the quality of the grassland was divided according to fractional vegetation cover (FVC). Currently, different standards exist for the FVC division of changing grasslands [11], and no specific specification exists for the threshold value of changing grasslands. Therefore, a relatively simple classification standard for grassland change was selected to achieve a unified division of grassland change under different natural conditions in the study area. The spatial pattern presented by the grassland quality division results was consistent with the theoretical spatial pattern based on the physical geographical environment of the study area and the spatial pattern known from actual investigations, including field surveys and interviews. This consistency further validates the reliability of the grassland data. Judging from the changes in the area proportion of FVC of different grades, the composition quality of grasslands has been continuously improving overall, which aligns with existing research indicating that vegetation NDVI and FVC have been improving in the Sanjiang Source area of the Qinghai–Tibet Plateau in recent years [5,54]. Additionally, the distribution results of middle and low-coverage grasslands were consistent with previous studies [4] on degraded grasslands, demonstrating the rationality and reliability of this research method and the results. However, the uncertainty in grassland coverage mainly depends on the NDVI value [55]. A more scientific and reliable remote sensing evaluation scheme for grassland quality warrants further discussion.

In terms of changes in grassland types, the type alterations (Table 5) were similar to existing research results [53], confirming that climate change was causing variations in land cover types. However, the results of C_v (Figure 8) also confirmed that human activities were a significant factor in the transformation of grassland types, such as overgrazing and urbanization construction. Although the intensity (Figure 5) and stability (Figure 8) of quality changes were high overall (Figure 6 and Figure 7), the spatial pattern was relatively pronounced. The intensity of quality change in the alpine region was high, and stability was poor. The fluctuation range of changing grasslands began to increase, indicating a shift in the stability of the grasslands. This also demonstrates that more specific areas of unstable grassland distribution can be identified through detailed spatial observation, which is lacking in the existing literature [14,15,16] we searched. Finally, it was found that changes in grassland stability were driven by climate change and human activities (Table 6). The influence of dual-factor coupling was stronger (Figure 9 and Figure 10). This result is similar to that of related studies [9]. Some studies have also suggested that human activities may exert more significant stress on the improvement of grassland quality, leading to an uncontrollable increase in the future stability trend of grassland [7].

Overall, the background factors determining the development environment of the grassland may dictate whether the grassland itself remains stable. The primary factors causing stable changes in grassland in high-cold regions may be climatic factors, whereas the stability of grassland in low-warm regions is mainly influenced by human activities. However, this study only realized the driving study of two-factor collaboration on grassland stability, and the driving effect of multi-factor collaboration on grassland stability will be an important research direction in the future. Even so, the more detailed detection and attribution analysis of alpine grassland ecosystem stability in this study provides a new perspective on the evolution of alpine grassland, which can offer an important reference for the Chinese government to implement ecological protection strategies in the context of global change.

5. Conclusions

The accuracy of regional grassland extraction, achieved by establishing a long-term series of stable samples and utilizing the random forest classification method, remained consistent and dependable. The spatial distribution of grasslands displayed a gradual transition from high-cover grasslands in the southeast to low-cover ones in the northwest. This trend is consistent with the predominant hydrothermal environment shaping the regional physical geographical spatial pattern.

Over the past 30 years, alongside changes in grassland types, the quality of grasslands in Guoluo has undergone alterations. Compared to 1990, there has been a slight decrease in the proportion of grassland area in Guoluo. Although overall grassland growth has seen gradual improvement, localized changes have been observed consistently.

Following 2005, the area experiencing changes in grassland quality exhibited significant fluctuations, suggesting a potential decline in ecosystem stability. However, despite these fluctuations, the results based on C_v indicate that the grassland ecosystem in Guoluo remains relatively stable. Only a few regions have experienced substantial changes attributed to climate change and human activity stressors. Climate change primarily affects high-cold regions, while human activities dominate low-warming areas.