Long-Term Spatiotemporal Changes and Geo-Information Tupu Characteristics of Qinling Mountains Ecosystem Pattern from 1986 to 2020

Highlights

What are the main findings?

• An optimized Support Vector Machine (SVM) classification method, based on historical field samples and multi-feature fusion, was developed. This method successfully achieved land cover classification and constructed a high-precision dataset for the period of 1986–2020 in the Qinling Mountains.
• Forests constituted the dominant land cover type, followed by croplands. From 1986 to 2020, the land cover pattern in the Qinling Mountains experienced substantial changes, accompanied by an overall improvement in the ecological environment.

What is the implication of the main finding?

• The high-precision long-term dataset and spatiotemporal flow analysis capabilities provide a robust technical framework and data foundation for systematically studying the historical evolution of regional ecological environments.
• Land cover in the Qingling ecological barrier has improved significantly over the past 34 years, a trend that appears linked to the implementation of targeted ecological restoration projects and government policies.

Abstract

The Qinling Mountains ecosystem serves as a vital ecological barrier and geographic demarcation line in China. Monitoring long-term land cover changes in the Qinling Mountains is essential for ecosystem pattern evaluation, environmental protection, and sustainable development. Focusing on the Qinling Mountains in Shaanxi Province, this study aimed to quantify the land cover changes from 1986 to 2020 using remote sensing and GIS technologies. An optimized Support Vector Machine (SVM) classification method was developed using Landsat satellite images and historical field samples. The method was employed to conduct land cover classification across eight discrete time periods: 1986, 1990, 1995, 2000, 2005, 2010, 2015, and 2020. The average overall accuracy (OA) of the classification results for the eight time periods was 96.42%, with a Kappa coefficient (K) of 0.9230, thus confirming the reliability of the mapping results. We subsequently developed a spatiotemporal Geo-information Tupu that facilitated a detailed analysis of land cover changes in the study area across different periods. The results show the following: (1) Forest was the dominant land cover type, followed by cropland. From 1986 to 2020, the forest, impervious surface, and water body areas showed overall increasing trends, although fluctuations were observed over time, and the increase was estimated at 6677.30 km², 557.57 km², and 135.71 km², respectively. In contrast, the areas of cropland, grassland, and bare soil showed a fluctuating decreasing trend, with a decrease in areal coverage of 2790.57 km², 1528.76 km², and 3042.66 km², respectively. During the study period, the forest area experienced the greatest increase but maintained the lowest dynamic degree. In contrast, bare soil showed the largest decrease and the highest dynamic degree. (2) A total of 30.74% of the area underwent dynamic changes during the study period, with the most active transformation occurring after 2010; these changes were mainly manifested in the outflow of cropland (4997.27 km²), the transfer of forest (8557.43 km²), and the expansion of impervious surfaces (771.33 km²). In conclusion, the overall ecological environment is improving. The results demonstrate a land cover reconstruction process that enables the management department to rationally utilize natural resources in the Qinling Mountains.

1. Introduction

Land cover plays an important role in natural resource management, ecological environment monitoring, climate change research, and related policy formulation [1], and land cover change directly reflects human–environment interactions, displays a great deal of diversity, and encompasses the entire Earth’s ecosystem [2]. Long-term land cover mapping can objectively and accurately reveal the pattern of land cover change, thus providing a scientific basis for understanding human–environment relationships and promoting integrated protection of mountains, water, forests, croplands, lakes, grasslands, etc. As the most important climatic boundaries that divide the North and South of China, the Qinling Mountains is a vital ecological security barrier, playing a key role in maintaining regional biodiversity, water resources, and climate stability. However, its land cover has experienced rapid changes since the 1980s due to human activities such as early deforestation, agricultural expansion, and urbanization, in addition to ecological degradation resulting from resource exploitation [3]. In recent years, the implementation of the Ecological Redline Policy and National Park Initiatives has aimed to enhance and maintain ecological security in the region. Therefore, there is an urgent need to examine the spatial distribution, attributes, and historical changes of regional land cover, which is essential for effective restoration of Ecological Space Governance in the Qinling Mountains.

In recent years, with the rapid development of remote sensing and Earth observation, as well as massive data processing technologies, land cover classification methods have evolved from manual interpretation [4], pixel-based shallow classification [5,6], and object-oriented classification [7,8] to the current stage of machine learning and deep learning methods [9,10,11]. Deep learning methods can solve the problems of shallow classification algorithms, which overly rely on manual feature annotation and have a poor generalization ability, but they require a large number of field samples and are more limited by the spatiotemporal accuracy of remote sensing classification data sources [12]. Support Vector Machine (SVM) is widely recognized as an efficient classical machine learning algorithm; it can effectively establish the coupling relationship between feature types and spectral information; it solves the problems of dimensional catastrophe, over-learning, and local minima in the classification method [13]; and it does not require large-scale and numerous samples for model parameter training and can handle complex and high-dimensional data that include many predictive variables and imbalanced samples [2,14,15]. Thus, it is widely used in land cover classification and dynamic monitoring with satisfactory classification accuracy [16,17,18,19,20]. On the basis of these classification methods, a series of algorithms, largely falling into two categories, (1) time series change detection and classification methods and (2) post-classification comparison methods, have been proposed and widely used for land cover time series classification [21]. The approaches in the first category generally adopt time series remote sensing images (e.g., Landsat TM/OLI, Sentinel, and MODIS), along with methods such as Dynamic Time Warping (DTW) and deep learning (CNN, RNN, LSTM, and Transformer-based models), and they have been shown to effectively improve the accuracy of time series classification [22,23,24,25]. However, they are not suitable for frequent disturbances or irregular observations, and they are not able to provide detailed change information. Although post-classification comparison methods are particularly time-consuming, they can achieve high classification accuracy with machine learning [21], but this technique ignores temporal dependencies and utilizes features independently from each other that can be derived from remote sensing time series [24]. Additionally, time series land cover change analysis mainly focuses on the spatiotemporal dynamics of land cover types. Its research content mainly includes the transformation of land cover types and changes in spatial distribution. Frequently utilized analytical methods are transition matrices, trend analysis, dynamic degree change assessments, and landscape index calculations. The Geo-information Tupu is an innovative approach that captures comprehensive data on spatial patterns and temporal characteristics of land cover change across multiple spatial and temporal scales using the graph unit; this method effectively illustrates the temporal evolution and spatial differentiation of natural processes and socio-economic development at the regional level [26]. Thus, this technology has been widely applied in the spatiotemporal change analysis of land cover [27,28].

A number of studies have been conducted focusing on land cover classification and the analysis of its change characteristics in the Qinling Mountains. Examples include land cover mapping using low- and medium-resolution remote sensing images combined with spectral and plant canopy structure features [29], manual and machine interpretation [30,31], and supervised and unsupervised classification [32]. Additionally, the spatiotemporal dynamics of land cover were analyzed, combined with the ecological environment quality index and cold/hot spot analysis [33]. This analysis revealed the transfer matrix and its correlation characteristics with ecological effects. However, the complexity of the surface characteristics of the Qinling Mountains makes field surveys difficult. Images are also affected by terrain, illumination, weather, and other conditions. It is less likely that field data are used in the classification [34]. While deep learning methods have been extensively researched and applied to date, they require large volumes of samples, and they are not suitable for long-term land cover classification in the Qinling Mountains, given the scarcity of field data. Instead, traditional machine learning methods that rely less on sample size may be more adaptable to this region. Moreover, the classification accuracy of machine learning-based methods is generally highly dependent on the selected features. Thus, during the classification process, it is necessary to optimize feature selection according to regional characteristics to improve classification accuracy. Additionally, most existing studies have utilized low- and medium-resolution satellite images and suffer from relatively narrow time series spans and compromised temporal continuity, thus failing to adequately reveal temporal patterns in land cover. Consequently, employing post-classification comparison methods combined with machine learning is necessary for long-term series land cover classification in this region, as this enables a comprehensive understanding of its spatiotemporal dynamic evolution.

The main objectives of this study were (1) to develop a historical field sample data-driven SVM classification method and construct a high-precision land cover dataset for the period of 1986–2020, based on long-term Landsat images, and (2) to construct a Geo-information Tupu to reveal the change process over multiple spatial and temporal scales, thereby intuitively reflecting the historical changes of land cover in the Qinling Mountains during the 34-year period. The results of this study will help to gain a deeper understanding of the spatial pattern of land cover in the region and the evolutionary processes while providing scientific support for the more effective implementation of ecological protection measures.

2. Study Area and Data

2.1. Study Areas

The study area refers to the narrow definition of the Qinling Mountains located in Shaanxi Province, which is representative of China’s north–south climate divide and serves as a strategically important ecological barrier. It connects Henan Province and Hubei Province in the east and Gansu Province in the west, stretching about 500 km. Bounded to the south by Sichuan Province, the boundary is defined by the northern bank of the Hanjiang River and the northern edge of the Hanzhong–Ankang Basins. Overlooking Guanzhong Plain in the north, the boundary is defined by the foothill line of the northern slope of the Qinling Mountains, with a width of 150 km. The total area is about 58,200 km², with the geographic coordinates ranging from 105°29′18″E–111°01′54″E to 32°28′53″N–34°32′33″N (Figure 1).

The Qinling Mountains are steep on the northern slope and gentle on the southern slope, with the highest elevation of 3771 m above sea level at the main peak of Mount Taibai and the lowest point of about 250 m at the bottom of the southern slope. There are differences in land cover between the northern and southern slopes. The main beam of the Qinling Mountains is the watershed between the Yellow River Basin and the Yangtze River Basin, the demarcation line between the warm temperate semi-moist monsoon climate zone and the northern subtropical humid climate zone. The complex topography and geomorphology have formed a rich and diverse mountain ecosystem. The administrative division encompasses the entire area of Shangluo municipality, as well as parts of the municipalities of Xi’an, Baoji, Weinan, Hanzhong, and Ankang, comprising 39 counties and districts in whole or in part, with a resident population of 4.89 million.

2.2. Data and Preprocessing

2.2.1. Satellite Images

Classified images were selected from the Landsat series, with a spatial resolution of 30 m acquired in eight time periods: 1986, 1990, 1995, 2000, 2005, 2010, 2015, and 2020. The Landsat5 TM (Thematic Mapper) sensor was used in 1986–1995, the Landsat5 TM and Landsat7 ETM+ (Enhanced Thematic Mapper Plus) sensors were used in 2000–2010, and the Landsat7 ETM+ and Landsat8 OLI (Operational Land Imager) sensors were used in the 2015–2020 period.

GEE (Google Earth Engine) is a cloud-based platform for geospatial analysis that can efficiently process satellite remote sensing images [35]. In this study, we used surface reflectance datasets from Landsat Collection 2 Level-2 available on GEE, which had already been calibrated and corrected for atmospheric effects and scan line striping. The atmospheric correction algorithm applied to Landsat 8 imagery is the Land Surface Reflectance Code (LaSRC), while the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) is used for atmospheric correction of Landsat 4–7 imagery. We further filtered the data based on the study area’s boundaries and focused on images from May to September, which is the main vegetation growing season. We further processed the data by selecting relevant bands, removing clouds, and applying percentile synthesis to generate historical time series datasets for eight time periods. These datasets effectively show how land cover types in the Qinling Mountains change over space and time.

2.2.2. Field Samples

In this study, we integrated three types of field-measured datasets: geographic national conditions monitoring [36], continuous forest resources inventory [37], and ground-based verification; all data types are point features. The integration resulted in sample datasets with an extended time series, a broad spatial coverage, diverse data types, a uniform distribution, and high accuracy. The measured data were collected concurrently with the images used for classification. The geographic national conditions monitoring dataset mainly uses remote sensing image interpretation samples, which take an area with a consistent time period, contiguous topography, and landscape and a minimum difference in geographic environment as the smallest sampling unit, including ground photographs, as well as corresponding land type information in the two periods of 2015 and 2020. The continuous forest resources inventory data mainly utilize the land type attributes of the sample points to facilitate the analysis of long-term sample migration from 1986 to 2014. Ground-based verification data were acquired in August 2020. To ensure strict temporal alignment with the classification tasks for each period, we selected ground-measured samples corresponding to the year of each remote sensing image period, and, based on these time-matched samples, we constructed training and validation sets specific to that epoch.

Based on the spectral and textural characteristics of the pixel information, region growing was used to gradually merge the similar pixels around the ground-measured sample points into the existing region to form the sample surface as the region of interest (ROI) in order to perform the classification task [38]. To ensure the validity and reliability of the ground-measured datasets, the ROIs were analyzed for stability and separation. In the stability screening process, ROIs were first selected to match the timeframes of the eight discrete classified images. Subsequently, only the samples that remained free of land cover changes between 1986 and 2020 were retained. On this basis, the Jeffries–Matusita distance [39] between each class in the sample datasets was calculated. After iterative optimization, the separation degree of these distances reached a value greater than 1.8. Finally, a total of 4730 ROIs were screened to build high-quality training and validation datasets. These included 2752 from geographic national condition monitoring, 1273 from continuous forest resources inventory, and 705 from ground-based verification. Then, we overlaid samples on images of all eight time periods to assess features, manually excluded excess forest and insignificant samples, and obtained the final sample points for each classified year (

Table 1. The number and source of sample points for each classified year.

Time Period	1986	1990	1995	2000	2005	2010	2015	2020
Training sample points	1956	1590	1874	1467	1826	1755	1961	2097
Validation sample points	838	681	803	629	783	752	840	899
Source	continuous forest resources inventory						geographic national conditions monitoring	geographic national conditions monitoring and ground-based verification

).

3. Methodology

We developed an optimized SVM classification method driven by historical field samples and multi-feature fusion, tailored to regional characteristics. The analysis employs a Geo-information Tupu and dual-dimensional indicators, specifically for time series and spatial transfer metrics, to examine changes in land cover types within the Qinling Mountains over a period of 34 years. Figure 2 presents the overall research framework. The classification system incorporates spatial features of the regional land surface, spectral properties from time series remote sensing data, and field-measured datasets. Land cover types are categorized into six classes: cropland, forest, grassland, impervious surface, bare soil, and water body.

3.1. SVM Classification Method

Eight time periods of reflectance images were used to classify historical land cover using the optimized SVM method. The field samples were divided into training samples and validation samples according to a ratio of 7:3. The model’s input features included raw spectral bands (Blue, Green, Red, NIR, SWIR1, and SWIR2), topographic variables (elevation, slope, and aspect), the Normalized Difference Vegetation Index (NDVI) [40], and the Normalized Difference Water Index (NDWI) [41]. The NDVI and NDWI values used in this study are multi-temporal composite features that integrate seasonal dynamics by averaging over the growing season. The SVM model incorporated a radial basis function (RBF) kernel to capture nonlinear features. Hyperparameters were systematically optimized using a grid search over an exponential sequence of values ($\mathrm{C}:2^{-5}\mathrm{to}{2}^{15}$; γ: $2^{-15}\mathrm{to}{2}^3$). A 5-fold cross-validation was employed to assess the OA for each parameter combination, and the ($\mathrm{C}$, γ) pair that yielded the highest mean cross-validation OA was selected to train the final model and evaluate its generalization capability. Ultimately, the model achieved precise classification of the six land cover types in the Qinling Mountains based on the optimal parameter and feature combinations. For each year, the classification accuracy was evaluated using OA, K, user’s accuracy (UA), and producer’s accuracy (PA) based on the confusion matrix [42].

3.2. Characterization Analysis of Spatiotemporal Evolution

The Geo-information Tupu is established based on raster pixels of land cover classification results from 1986 to 2020 [43]. The term “Tu” of the spatial scale characterizes different land cover types, defined as P ($P_1,P_2\dots P_n$), and “Pu” of the time scale characterizes different time points, defined as T ($T_1,T_2\dots T_m$). “Tu” and “Pu” features are synthesized to form spatial-attribute-process-integrated data to construct a spatiotemporal evolution knowledge graph [26]. Combined with the land cover and its historical development characteristics in the Qinling Mountains, this study determined the “Tu” unit as 6 (n = 6) and the “Pu” unit as 4 (m = 4), and four periods with typical representativeness, namely, 1986, 2000, 2010, and 2020, were selected. The specific calculation formula for the graph unit is as follows:

$$TP={1000\times P}_{T1}+{100\times P}_{T2}+{10\times P}_{T3}+P_{T4}$$

(1)

where $TP$ represents the graph unit, which is a 4-digit graph coding. $P_{c1}$, $P_{c2}$, $P_{c3}$, and $P_{c4}$ represent the land cover types of the four typical periods (1986, 2000, 2010, and 2020, respectively). Based on the land cover data for the four periods, the final 4-digit graph coding is obtained using the raster calculator in GIS software. Then, the Geo-information Tupu is divided into six types: stable and unchanging, early-stage change, mid-term change, late-stage change, repeated change, and continuous change. Its mapping relational expression and the significance of its characterization are shown in

Table 2. Geo-information Tupu relationship expression.

Types of Geo-Information Tupu	Graph Relational Expression	Significance of the Graph
Stable and unchanging	$P_{T_1}$ = $P_{T_2}$ = $P_{T_3}$ = $P_{T_4}$	Areas where the land cover type has remained unchanged
Early-stage change	$P_{T_1}$ ≠ $P_{T_2}$ = $P_{T_3}$ = $P_{T_4}$	Changes occurred in 1986–2000
Mid-term change	$P_{T_1}$ = $P_{T_2}$ ≠ $P_{T_3}$ = $P_{T_4}$	Changes occurred in 2000–2010
Late-stage change	$P_{T_1}$ = $P_{T_2}$ = $P_{T_3}$ ≠ $P_{T_4}$	Changes occurred in 2010–2020
Repeated change	$P_{T_1}$ ≠ $P_{T_2}$ = $P_{T_3}$ ≠ $P_{T_4}$ or $P_{T_1}$ = $P_{T_2}$ ≠ $P_{T_3}$ = $P_{T_4}$ or $P_{T_1}$ ≠ $P_{T_2}$ ≠ $P_{T_3}$ = $P_{T_4}$, also $P_{T_1}$ = $P_{T_4}$	The number of changes in land cover type from 1986 to 2020 is greater than 1 and is also the same for both the 1986 and 2020 periods.
Continuous change	$P_{T_1}$ ≠ $P_{T_2}$ = $P_{T_3}$ ≠ $P_{T_4}$ or $P_{T_1}$ = $P_{T_2}$ ≠ $P_{T_3}$ = $P_{T_4}$ or $P_{T_1}$ ≠ $P_{T_2}$ ≠ $P_{T_3}$ = $P_{T_4}$, also $P_{T_1}$ ≠ $P_{T_4}$	The number of changes in land cover type from 1986 to 2020 is greater than 1, and the types are not the same for the 1986 and 2020 periods.

.

Five indicators, namely, the amount of land cover area change, the average annual increment, single land cover dynamic degree $K$ [44], integrated land cover dynamic degree $L{C}_s$ [45], and the Human Disturbance Index $L{C}_f$ [46], were used to quantitatively express the change dynamics of land cover in the spatiotemporal domain. Here, $K$ is highly sensitive to the initial area and reflects the relative rate of change. It also changes linearly in direct proportion to the increase or decrease in the area of land cover types during the study period. The calculation formula is as follows:

$$K={\displaystyle \frac{U_2-U_1}{U_2}}\times {\displaystyle \frac{1}{T}}\times 100\%$$

(2)

where $K$ is the land cover dynamic degree during the study period; $U_1$ and $U_2$ represent the areas of the specific land cover type at the beginning and end of the study period, respectively; and $T$ is the time interval.

$\mathrm{L}{\mathrm{C}}_{\mathrm{s}}$ can quantify the change rate and effectively reflect the overall trend of land type changes over the monitoring period. The calculation formula is as follows:

$$L{C}_s={\displaystyle \frac{\sum _{i=1}^n∆L{U}_{i-j}}{2\sum _{i=1}^{n}L{U}_i}}\times {\displaystyle \frac{1}{T}}\times 100\%$$

(3)

where $L{U}_i$ represents the area of land cover type $i$ at the initial stage of monitoring, $∆L{U}_{i-j}$ denotes the absolute value of the area of land cover type $i$ converted to non-$i$ land cover types during the monitoring period, and $n$ stands for the total number of land cover categories. Higher $L{C}_s$ values reflect more frequent land cover transitions and more intense land cover change in the region, whereas lower $L{C}_s$ values indicate greater land cover stability and negligible changes.

$L{C}_f$ can quantify the impact of human activities on the ecological spatial pattern of the Qinling Mountains. The calculation formula is as follows:

$$L{C}_f={\sum }_{i=1}^n{\displaystyle \frac{A_i\times P_i}{A_a}}$$

(4)

where $A_i$ represents the area of different land cover types, $P_i$ represents the intensity of impacts of land cover types, $A_a$ is the total area of the study area, and $n$ stands for the total number of land cover types. Larger $L{C}_f$ values correspond to more frequent human activities, higher impacts on natural ecological spaces, and greater potential for ecological and environmental problems in the region; smaller $L{C}_f$ values, by contrast, indicate the opposite.

Referring to the Blue Book of China’s Geographic National Conditions in 2017, the intensity of impacts of different land cover types in the Qinling Mountains was determined, with values ranging from 0 to 1, as shown in

Table 3. Intensity of impacts of different land cover types.

Land Cover Types	Cropland	Forest	Grassland	Impervious Surface	Bare Soil	Water Body
Intensity of impacts	0.55	0.1	0.23	0.95	0.23	0.115

. Higher values mean more human activity, and lower values indicate more natural land. Based on the transfer matrix, the evolution process in the spatial domain was clarified by combining the graphical characteristics of land cover types over time.

4. Results

4.1. Land Cover Classification Information Extraction

Land cover classification maps of the Qinling Mountains in eight time periods during 1986–2020 are shown in Figure 3. The figure shows that the Qinling Mountains are mainly covered by forest, with an area share above 70% during the study period. Cropland accounts for about 10%, primarily located in the rural and suburban areas of Shangluo and Ankang municipalities, as well as the eastern part of Xi’an City. Impervious surfaces, grassland, bare soil, and water bodies together make up a small proportion. Overall, the region is dominated by forest cover, with rivers, lakes, and cultivated land or man-made surfaces scattered in valleys and gentle slopes.

Table 4. Evaluation results of classification accuracy for eight time periods (unit: %).

Land Cover Types	1985		1990		1995		2000		2005		2010		2015		2020
	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA	UA	PA
Cropland	81.54	87.83	96.59	76.52	92.57	94.09	99.11	99.18	89.07	96.64	91.29	95.07	94.43	90.37	90.34	90.29
Forest	99.84	97.68	99.96	99.54	99.93	99.91	99.50	92.16	99.88	99.94	99.96	99.93	99.85	99.62	99.83	99.43
Grassland	78.07	92.41	91.78	97.10	45.36	87.59	48.68	61.49	45.23	93.94	45.87	92.65	52.04	78.85	55.89	58.03
Impervious Surface	82.24	76.73	82.24	94.99	93.12	99.05	83.01	83.01	99.82	98.58	97.85	93.54	100	99.37	95.17	93.36
Bare Soil	86.41	61.81	86.41	76.72	92.40	55.73	88.68	92.87	97.80	68.99	98.59	86.40	91.44	90.40	72.72	72.39
Water Body	90.39	97.35	90.39	97.35	98.90	96.26	97.56	98.85	99.07	100	100	100	99.35	100	95.48	94.27

presents the accuracy of each land cover classification for eight time periods, as evaluated by OA, K, UA, and PA. Overall, the classification accuracy is high, although notable variations exist across different land cover types and years. UA and PA for forest and water bodies remain consistently high and stable over time, whereas cropland and impervious surfaces exhibit minor temporal fluctuations. Grassland and bare soil present greater classification difficulties, with lower accuracy in specific years.

Furthermore, we compared the optimized SVM classification results with those of existing mainstream land cover products, namely, GLC_FCS30 [47], which has a spatial resolution of 30 m and covers eight time periods (1986, 1990, 1995, 2000, 2005, 2010, 2015, 2020). The evaluation results are shown in

Table 5. Comparison of classification accuracy.

Time Period	Optimized SVM Method		GLC_FCS30
	OA (%)	K	OA (%)	K
1986	91.74	0.8723	83.68	0.7344
1990	95.11	0.8579	90.26	0.7073
1995	98.23	0.9305	88.00	0.8060
2000	95.48	0.9400	79.05	0.7100
2005	98.27	0.9593	87.65	0.8031
2010	98.03	0.9585	92.40	0.8383
2015	97.94	0.9575	87.95	0.7696
2020	96.57	0.9082	86.42	0.6787
Average	96.42	0.9230	86.93	0.7559

. Compared with this classification, the accuracy of the optimized SVM method was satisfactory, with the OA of the classification result for each time series reaching over 90%. This fully demonstrates the effectiveness of field measurement information for remote sensing classification models. It provides data support for the accurate analysis of the spatiotemporal evolution characteristics of land cover in the Qinling Mountains.

4.2. Analysis of the Spatiotemporal Evolution of Land Cover

4.2.1. Characteristics of Temporal Changes

Between 1986 and 2020, the characteristics of different land cover types in the Qinling Mountains underwent changes over time (Figure 4). Forest showed a growth trend, with a total increase of 6677.302 km² from 1986 to 2020, and the area peaked in 2020, with an average annual increase of 196.39 km², at a rate of 15.42%, and a dynamic degree of 0.39%. The growth of forest land accelerated in the later stage of the monitoring period, with the area increasing by 4189.93 km², of which the growth rate was the highest from 2010 to 2015, with a dynamic degree of 1.74%, and the growth rate slowed down from 2015 to 2020, with the area increasing by 78.069 km² and a dynamic degree of 0.16%. Cropland in general had a trend of “fluctuation in the early stage and continuous decrease in the middle and late stage”. In 1986, cropland area was the largest (8022.32 km²), and during the 34-year period, the area reduced by a total of 2790.568 km², with an average annual decrease of 82.08 km². The overall decrease rate was 34.78%, and the dynamic degree rate of change was −1.57%. In the late stage, cropland had the highest vulnerability to change, and the shrinkage was significantly higher than that in both the early and middle stages.

Grassland overall followed a “growth (1986–1995)–plummet (1995–2000)–rise (2000–2010)–decline (2010–2020)” trend. In 1995, the grassland area was the largest, at 4899.466 km², and then it decreased by 1528.76 km² during the 34-year period, with an average annual decrease of 44.96 km², an overall decrease of 46.98%, and a dynamic degree of −2.61%. The greatest change in grassland area occurred between 1995 and 2000, with an average annual decrease of 496.52 km² and a dynamic degree of −13.18%, and the forest area increased in this period. In the middle of the monitoring period, the increase in grassland area was accompanied by a continuous decrease in cropland. The impervious surfaces continued to expand, and the growth rate gradually accelerated, increasing by a total of 557.566 km² from 1986 to 2020, with the largest area of 813.61 km² in 2020. The average annual increase was 12.3 km², the overall increase was 217.76%, and the overall dynamic degree was 2.02%. The growth rate of impervious surfaces was slow in the early stage, with an average annual growth of about 12.3 km² and a dynamic degree of 1.80%. Bare soil showed a continuous decreasing trend, with a total decrease of 3042.662 km² from 1986 to 2020, with an average annual decrease of 89.49 km², a decrease of 90.82%, and an overall dynamic degree of −29.08%. Furthermore, 1986–2005 was a period of rapid decrease, among which the area from 2000 to 2005 decreased the most, with an average annual decrease of 301.32 km² and a dynamic degree of −17.55%. After the year of 2005, the bare soil area slowly declined and tended to stabilize. Water bodies showed a trend of “fluctuating and rising in the early and middle periods, and growing steadily in the later periods”. The water body area was the largest in 2020, increasing by 135.709 km² compared with that in 1986, with an average annual increase of 3.99 km², which is an increase of 154.84%, and the dynamic degree was 1.79%. In 2005, there was a higher increase in the water body areas, with an average annual increase of 14.15 km², and the dynamic degree was 10.88%.

Overall, cropland, bare soil, and grassland each exhibited a declining trend, with bare soil experiencing the most pronounced change and the largest decrease in area. In contrast, forest, impervious surfaces, and water bodies generally increased, with impervious surfaces showing the highest rate of expansion and forest having the smallest rate of increase. The sequence of $L{C}_f$ of integrated land cover dynamics for the entire study area was as follows: late (1.08%), early (0.96%), and middle (0.93%). The Human Disturbance Index at four key time points of 1986, 2000, 2010, and 2020 was calculated as 0.180, 0.176, 0.172, and 0.157, respectively, indicating that the influence of human activities on the whole Qinling Mountains ecosystem remains relatively low and is steadily decreasing.

4.2.2. Space Transfer of Geo-Information Tupu

(1)

Characterization of overall change graph

Based on the Geo-information Tupu, 1239 graph units were formed by using GIS spatial analysis methods to construct a knowledge graph of the spatiotemporal evolution of land cover types in the Qinling Mountains from 1986 to 2020. This approach enabled multiple expressions of time-domain information at the spatial granularity and transferred temporal changes from the image meta-spectrum to the spatial graph. The spatial distribution of different Geo-information Tupu types is shown in Figure 5. As can be seen in the figure, during the study period, a total of 69.26% of the area remained unchanged, mainly distributed within the ecological red line, Qinling National Park, and the key and core protection areas of the Qinling Mountains. A total of 30.74% of the area showed a dynamic trend of change. These changes were mainly distributed outside the core ecological protection areas of the Qinling Mountains. There was a relatively high concentration of dynamic areas along rivers and urban fringes of towns and cities. The Jialing River and Han River watersheds in the southwestern Qinling Mountains also showed a clear trend of dynamic land cover changes over time.

Among the changed areas, the repeated and continuous change types of Geo-information Tupu are larger, accounting for 57.22% of the total change area. The area of early-stage changes accounts for 21.73%, concentrating in the eastern part of the Qinling Mountains of Hanzhong City, as well as in the area along the northern foothills of the Qinling Mountains. The mid-term changes account for 4.27% of the area, mainly occurring along the Danjiang River and the Hanjiang River Basin; the late-stage changes account for 16.77%, mainly distributed in the basins of the Jialing River and the Hanjiang River at the southern foot of the Qinling Mountains, as well as in areas along the northern foot of the Li Mountain range where human activities are more intense.

(2)

Characterization of typical land cover type change Tupu

In this study, a transfer matrix is used to quantitatively characterize the spatial changes of six types of “Tu” units in four “Pu” units; the transfer matrices are shown in Figure 6. Figure 7 shows a spatial overlay analysis of increased and decreased land cover types from 1986 to 2020. The spatial distribution of typical land cover types in terms of detailed transfer-out and transfer-in is shown in Figure 8. The statistical information of transfer areas and the main Tupu characteristics of different land cover types at various time dimensions are shown in

Table 6. Statistics on the transfer area and Geo-information Tupu characteristics of various types.

Land Cover Type	Stable and Unchanging	Repeated Change		Early-Stage Change				Mid-Term Change				Late-Stage Change				Continuous Change
				Transfer-Out		Transfer-In		Transfer-Out		Transfer-In		Transfer-Out		Transfer-In		Transfer-Out		Transfer-In
	Area	Area	TP	Area	TP	Area	TP	Area	TP	Area	TP	Area	TP	Area	TP	Area	TP	Area	TP
Cropland	2112.70	900.49	1511	827.46	1222	535.92	2111	424.79	1122	72.09	2211	1856.59	1112	593.02	2221	1888.44	1532	1012.16	2121
Forest	38,084.76	3300.66	2122	233.39	2111	3222.12	5222	96.13	2211	506.23	1122	878.98	2221	1391.92	1112	685.39	2121	3437.16	3232
Grassland	59.79	434.38	3533	876.31	3222	50.86	2333	168.64	3322	74.33	2233	192.21	3332	526.93	2223	1521.98	3232	576.94	5223
Impervious Surface	27.41	14.02	4114	57.79	4222	51.97	1444	4.23	4411	85.08	1144	13.64	4442	332.64	1114	138.61	4232	301.65	2114
Bare Soil	12.38	16.91	5115	1892.38	5222	1.32	2555	70.54	5522	10.44	3355	56.98	5552	107.75	1115	1299.60	5232	158.47	1445
Water Body	33.56	18.07	6166	3.19	6222	28.33	1666	0.76	6644	16.92	1166	3.65	6665	49.81	1116	24.68	6112	72.34	1446

Note: The unit of area is km²; the code 1 in the graph unit represents cropland, 2 represents forest, 3 represents grassland, 4 represents impervious land, 5 represents bare soil, and 6 represents water body.

.

From 1986 to 2020, the total area where land cover types were transformed within the study area is 13,216.370 km². Among them, the area of cropland transfer-out is 4997.270 km², being mainly transformed into forest (3638.525 km²), grassland (561.511 km²), and impervious surfaces (514.574 km²), and the area transformed into forest land accounts for 72.81% of the total area of cropland outflow. The spatial location of transferred cropland is mainly distributed in the gently sloping valleys of the Han River and Jialing River Basins in the southern foothills of the Qinling Mountains, the shallow hills of the Li Mountains in the northern foothills of the Qinling Mountains, and most of the areas in Shangluo City. The main types of Geo-information Tupu of cropland outflow are late-stage change and continuous change, corresponding to the graph units of “cropland–cropland–cropland–forest” and “cropland–bare soil–grassland-forest”, respectively, which account for a large proportion of the outflow area between 2010 and 2020. The unchanged area of cropland is 3013.191 km², which can be seen by combining the characteristics of the graph, of which 2112.702 km² are stable and unchanged, and 900.489 km² are transformed to other land cover types first and then restored to cropland in the study period. The area converted to cropland from other land cover types is 2213.190 km², the main sources of which are forest land and grassland, and the graph features a discrete distribution pattern, with changes mainly occurring in the middle and late periods.

The forest outflow area of 1893.889 km² was mainly transformed into cropland (1136.583 km²), grassland (523.609 km²), and impervious surfaces (136.075 km²). The transfer-out of forests mainly occurred between 2010 and 2020, with the main graph unit being “forest–forest–forest–cropland,” and the area converted to cropland accounted for 60.01% of the total transfer-out area. The area of unchanged forest was 41,385.42 km². Combined with the map characteristics, this indicates that 38,084.76 km² of forest remained stable and unchanged. Additionally, 3300.66 km² of forest was first transformed into other land cover types and then reinstated to forest during the study period, with the main graph unit being “forest–cropland–forest–forest”. The forest transfer-in area was 8557.430 km², and the main sources were bare soil and cropland. The spatial location was mainly distributed in the southwestern corner of the Qinling Mountains in the administrative area of Hanzhong City and Weinan Municipal District in the northern foothills, as well as in several areas in the southeastern corner of the study area. The main types of spatiotemporal graphs were early-stage change (with the main graph unit being “bare soil–forest–forest–forest”) and late-stage change (with the main graph unit being “cropland–cropland–cropland–forest”, accounting for 86.94% of the total area of late-stage change), while the mid-term change was rendered insignificant.

Changes in the impervious surfaces were mainly reflected in the transfer of other land cover types, with an area of 771.334 km², the main sources of which were cropland (514.574 km²), forest (136.075 km²), and grassland (74.518 km²). Most new impervious surfaces were found in towns and villages with high human activity, particularly around the outskirts of Ankang and Shangluo and near major transportation routes. The main change type of the spatiotemporal graph of impervious surfaces was late-stage change, with a total area of 332.644 km², and the main graph unit was “cropland–cropland–cropland–impervious surface,” accounting for 89.97% of the total late-stage change area. The transfer-in area was only 217.273 km², and the main transferred types were forest and cropland.

Bare soil and grassland mainly showed transfer-out changes, and the outflow area was 3319.502 km² and 2759.152 km², respectively. The main transfer types were forest and cropland, with a larger proportion of early-stage change and continuous change. The changes in the water body were reflected in the transfer of other land types, mainly cropland, accounting for 79.75% of the total transferred area. The changes mainly occurred in the later stages of the study period and were primarily located in the vicinity of rivers.

5. Discussion

5.1. Evaluation of Classification Results and Limitations of the Method

The land cover mapping approach presented in this study incorporates multi-source historical field samples into the classification process, leading to a significant improvement in classification accuracy. By integrating the analysis of spatiotemporal Geo-information Tupu characteristics, this method enables more effective tracking of ongoing changes than traditional transition matrices. Additionally, it identifies subtle temporal patterns and associates classification outcomes with real-world factors, such as policy changes in the Qinling Mountains. To validate the superiority of the optimized SVM method, a comprehensive comparison was conducted between the classification results of this method and those of mainstream existing land cover products. Compared with the widely used land cover product GLC_FCS30, the proposed method achieves an average accuracy increase of 9.49%. These results demonstrate the improved capability of the proposed approach to capture land cover characteristics. In addition to accuracy quantification, a comparative analysis of land cover spatial distribution across the GLC_FCS30 product was performed, as shown in Figure 9. This analysis revealed specific limitations of large-scale land cover products in local-scale applications. These products are typically evaluated using global average accuracy metrics, which do not effectively reflect their applicability to specific local areas or capture detailed spatial information. This problem is obvious in complex mountainous regions with few field samples. For instance, in the Qinling Mountains, some grassland areas were wrongly classified as forest.

However, the proposed method still has some limitations. Taking the classification results of 2020 as an example, the OA reached 96.57%, but the confusion matrix revealed two main issues. First, although historical field samples address the challenge of limited feature reuse in remote sensing images across extended time periods and varying climates, these samples are primarily from forested areas, representing 78.20% of the total. Second, the classification model works accurately for stable land cover types like forests and water bodies. However, it still needs improvement for grassland, bare soil, and cropland to lower the rate of misclassification. The cropland in seeding or fallow periods exhibits spectral and textural characteristics similar to grassland and bare soil, resulting in classification confusion. Future research should optimize sampling strategies by employing multi-phenological composites to enhance cropland representation and by supplementing bare soil samples collected during periods of vegetation growth. Additionally, introducing multi-texture features can help develop discrimination rules for confusable land cover types. Integrating multi-source remote sensing data leverages complementary information. All these strategies are expected to enhance classification accuracy.

5.2. Evaluation of Regional Characteristics of Land Cover Dynamics

Between 1986 and 2020, 30.74% of the land cover types in the study area underwent dynamic changes, associated with a combination of factors, including policy reforms, socioeconomic activities, and climate change. The period from 1995 to 2000 marks a watershed in the ecological protection of the environment and the development of natural resources in the Qinling Mountains. In 1998, the Shaanxi Provincial Government issued the “Order on the Implementation of the Natural Forest Conservation Project to Immediately Halt Harvesting of Natural Forests”, and a comprehensive logging ban was imposed in the Qinling Mountains; the recovery of forest coverage in this area may be related to the implementation of these policies. The cropland area increased between 1990 and 2000. This period coincided with major land management directives, including the 1994 “Regulations on Basic Farmland Protection” and the 1997 “Balance of Occupation and Compensation” policy. Since 2000, however, the area of cropland has continued to decrease; this change correlates with the 1999 initiation of the Grain-for-Green Project (GFGP) pilot in the Qinling Mountains. Variations in grassland often occur alongside changes in cropland and forest cover, which are likely influenced by land use policies, thus illustrating the impact of such policies on the regulation of the Qinling Mountains ecosystem. Overall, the ecological environment of the Qinling Mountains has continued to improve, which is largely consistent with previous studies [3,48,49]. However, vegetation cover changes induced by land cover change did not improve immediately after afforestation activities but rather increased a few years later [3].

Since 2000, urban infrastructure in cities, districts, and counties within the Qinling Mountains has expanded rapidly. This growth was associated with socioeconomic development and major projects such as the Xi-Han Expressway and the Xi-Cheng High-Speed Railway. These human activities mainly took place in low-mountain areas, including piedmont alluvial plains and flood plains [50]. As a result, natural surfaces such as cropland and forest were occupied. This is consistent with the findings of Cui [33] and Wang [51]. However, overall, the intensity of human impact on the Qinling Mountains ecosystem remains relatively low and has shown a consistent downward trend [52]. Moreover, following 2005, the Qingling Mountains, designated as one of the initial pilot areas for national ecological function zones, experienced increasing robust ecological protection policies. Then, the damaged land surface was gradually restored to natural vegetation, and the bare soil area gradually declined and stabilized. There was a higher increase in water body areas, which is likely attributable to the construction of projects such as “the Han-to-Wei River water diversion project” and “the Yinghu Wetland Park”. Following the implementation of ecological water replenishment, forest conservation, and source water protection, the water conservation capacity of the Qinling Mountains showed a steady increase [53]. This increase was associated with a significant expansion of the region’s water bodies during the later period. We want to clarify that our assessment of these changes as outcomes of ecological policies is qualitative, as we did not quantify the influence of climate variables. In future research, we aim to use causal inference models to distinguish policy effects from climate effects, thereby enhancing our research findings.

The Qinling Mountains represent the ancestral origin of the Chinese nation and serve as a significant symbol of Chinese culture. This study provides precise data on regional land cover and its complex evolutionary processes. From a historical perspective, the findings comprehensively illustrate the spatial and temporal patterns of land cover change in the Qinling Mountains, shaped by policy, natural factors, and human adaptation.

6. Conclusions

This study utilized eight discrete time periods of Landsat satellite images. The spectral bands, topographic variables, NDVI, and NDWI were integrated to develop an SVM-optimized classification method. This approach, supported by historical field samples, enabled the extraction of land cover classification information for the Qinling Mountains from 1986 to 2020. A Geo-information Tupu was constructed from the classification maps to show land cover dynamics in the Qinling Mountains over the past 34 years. This multilevel analysis can help optimize land use, guide ecological policy, and support regional sustainability. The main conclusions are as follows:

(1)

A real field samples-driven SVM classification model method was developed in this study. Using eight sets of ground samples collected at different times, a reliable and consistent land cover dataset covering 34 years was created for the Qinling Mountains. The method achieved an average accuracy of 96.42% and a Kappa coefficient of 0.9230.

(2)

Forest is the predominant land cover in the Qinling Mountains, followed by cropland. From 1986 to 2020, the fluctuating growth trend of the forest was evident, increasing from 43,291.693 km² to 49,968.995 km², with a slight decrease in 1995. The cropland area declined from 8022.317 km² in 1986, with a significant reduction after 2010, totaling a decrease of 2790.568 km². The grassland and bare soil areas also decreased, by 1528.759 km² and 3042.662 km², respectively. Impervious surfaces and water bodies expanded by 557.566 km² and 135.709 km². Forest had the largest area increase with the lowest dynamic degree, while bare soil had the largest decrease and the highest dynamic degree.

(3)

During the study period, a total of 30.74% of the land cover exhibited a dynamic changing trend. This change was mainly reflected in the transfer-out of cropland, with the typical spatiotemporal graph unit being “cropland–cropland–cropland–forest” and a total outflow area of 4997.270 km². The transfer-in of forested land was also significant, with the typical unit being “bare soil–forest–forest–forest” and a total inflow area of 8557.430 km². Additionally, impervious surfaces expanded mainly at the expense of cropland and forest, covering an area of 771.334 km². Overall, the changing trend was most significant after 2010.

From 1986 to 2020, land cover changes in the Qinling Mountains were characterized by accelerated ecological restoration and reduced human interference. Future management should balance ecological conservation, food security, and moderate development to ensure the long-term sustainability of the Qinling Mountains.