Remote Sensing Detection of Forest Changes in the South Ridge Corridor and an Attribution Analysis

Abstract

As the largest mountain range in Southern China, the natural vegetation of Nanling plays an irreplaceable role in maintaining the stability of the ecosystem and exerting its functions. The forested area of the Nanling Corridor encompasses 168,633 km², with a forest coverage rate exceeding 60% of all cities together. Long-term analysis of the temporal and spatial evolution of this forest and the disturbance factors in this region is of great importance for realizing the “dual carbon” goals, sustainable forest management, and protecting biodiversity. In this study, remote sensing images from a Landsat time series with a resolution of 30 m were obtained from the GEE (Google Earth Engine) cloud processing platform, and forest disturbance data were obtained using the LandTrendr algorithm. Using a machine learning random forest algorithm, the forest disturbance status and disturbance factors were explored from 2001 to 2020. The results show that the estimated disturbed forest area from 2001 to 2020 was 11,904.3 km², accounting for 7.06% of the total area of the 11 cities in the Nanling Corridor, and the average annual disturbed area was 595.22 km². From 2001 to 2016, the overall disturbed area increased, reaching a peak value of 1553.36 km² in 2008, with a low value of 37.71 km² in 2002. After 2016, the disturbed area showed a downward trend. In this study, an attribution analysis of forest disturbance factors was carried out. The results showed that the overall accuracy of forest disturbance factor attribution was as high as 82.48%, and the Kappa coefficient was 0.70. Among the disturbance factors, deforestation factors accounted for 58.45% of the total area of forest disturbance, followed by fire factors (28.69%) and building or road factors (12.85%). The regional distribution of each factor also had significant characteristics, and the Cutdown factors were mostly distributed in the lower elevations of the mountain margin, with most of them distributed in sheets. The fire factors were spatially distributed in the center of the mountains, and their distribution was loose. Building or road factors were mostly distributed in clusters or lines. These research results are expected to provide technical and data support for the study of the large-scale spatiotemporal evolution of forests and its driving mechanisms.

1. Introduction

Forests, as an indispensable part of the ecosystem, play an important role in soil and water conservation, climate regulation, ecosystem stabilization [1], and regulating the global carbon and water cycles [2]. Because of their function of absorbing carbon dioxide and releasing oxygen, forests are called a “natural oxygen bar” and the “lungs of the Earth”. As one of the most important types of vegetation on the Earth, forests are not only an important part of the biosphere but also the main bodies of terrestrial ecosystems and an indispensable part of the global carbon cycle [3]. Forests play a crucial ecological role in human health and sustainable development globally, including protecting soil and water resources, controlling wind and sand, provisioning timber and energy, and maintaining ecosystem stability [4], and they play a pivotal role in the mitigation of climate change, carbon sequestration, oxygen production, and the renewable use of resources [5]. The important role played by forests in the sustainable development of ecosystems and human societies has brought attention to the detection of changes in forest areas due to anthropogenic and natural causes [6]. Nanling is located in the ecological intersection zone of China’s subtropical zone, which is the transition zone of China’s northern and southern biota, with diverse vegetation types, obvious ecological gradients, and high sensitivity to environmental changes [7]. In addition, this region is an important biodiversity hotspot in China, covering a wealth of plant and animal species, and its forest ecosystem plays an important role in maintaining regional ecological balance and providing ecological services. In recent years, the forest ecosystem in the Nanling region has been disturbed to varying degrees due to the combined effects of natural disturbances and human activities [8]. Therefore, the study of forest disturbance in this region not only helps to reveal the response mechanism of subtropical forest ecosystems to disturbances but also provides a scientific basis for regional ecological restoration and management.

Monitoring forest resources and vegetation change has always been a very important task in forestry surveys, and a very important step is surveying forest resources. China conducts a forest resource survey every five years. However, due to the shortcomings of the traditional field survey, with high costs, a long period, low efficiency, and poor timeliness, it is difficult to monitor sporadic and short-term vegetation disturbance conditions [9]. Therefore, the detection of vegetation changes and the development of forestry planning on a large scale over a long period cannot be carried out based solely on field inspection. With the maturity of satellite remote sensing and computer technology, remote sensing has begun to play an increasingly important role in vegetation change detection. Satellite remote sensing data have the advantages of convenience, wide coverage, and timeliness and have gradually become the main means of vegetation change monitoring [10]. At present, many remote sensing platforms are used, such as SPOT, Landsat, Quickbird, Worldview, IKONOS, etc., among which Landsat is widely used to study the disturbance and recovery of forest vegetation [11]. There are studies utilized global 1990–1997 tropical moist forest cover data obtained from AVHRR data and used a stratified statistical sampling method to obtain a map of the deforestation area in each region from high-resolution imagery in 1990 and 1997 [12]. There are studies used the 2000–2005 Landsat remote sensing data as training data, extracted the global forest cover change from 1990 to 2000 globally, and evaluated it using U.S. regional samples. They obtained 93% accuracy for forest cover and 84% accuracy for forest cover change, while the global accuracy evaluation showed 88% accuracy for forest cover change [13].

Annual-scale algorithms for forest disturbance detection include LandTrendr [14], which uses a single spectral index, NBR, as an input detection metric, and currently has limitations such as delaying the detection of forest disturbances occurring in the fall and winter seasons by at least one year by using only Landsat synthetic observations each summer and detecting all forest disturbances simultaneously using a single spectral index or band. The Vegetation Change Tracking (VCT) algorithm utilizes a composite forest score detection metric, which currently suffers from problems such as difficulty in identifying some low-density forest disturbances such as forest thinning and selective logging [15]. The seasonality and trend breakpoint algorithm (BFAST) [16], which utilizes a single spectral index, NDVI, as an input detection metric, currently suffers from problems such as not being able to provide forest disturbance types. The sky-scale algorithm for forest disturbance detection includes the Continuous Change Detection and Classification (CCDC) algorithm, which utilizes the six spectral bands of blue, green, red, near-infrared (NIR), short-wave 1 (SWIR1), and short-wave 2 (SWIR2) in the Landsat imagery as the input detection indexes, and the algorithm has not yet been applied to detect low-density forest disturbance with weak spectral changes. At present, the algorithm still has problems such as lower detection accuracy of low-density forest disturbances with weak spectral changes. The Continuous Surface Disturbance Monitoring (COLD) algorithm [17], which utilizes the six bands of blue, green, red, NIR, SWIR1, and SWIR2 in Landsat imagery, as well as multiple metrics such as the probability of change in the chi-square distribution, the number of consecutive anomalous observations, and the angle of change as inputs, also suffers from the problem of requiring a large amount of data storage space with high computational costs.

The LandTrendr algorithm uses the observation data synthesized by Landsat every summer as the annual spectral value, constructs a time series for each pixel with a single optical spectral index or band as the input, and iteratively identifies a set of breakpoints based on regression and then fits the line segments between them [18]. The result of this process is an annual time series represented by segments, based on which a lot of information can be derived, such as the number of segments, length (in years), slope, and year corresponding to the vertex of the segment. The LandTrendr algorithm based on Landsat time-series data is currently one of the most widely used algorithms for forest disturbance detection, which can run efficiently and compute in parallel on the GEE cloud platform [19]. In addition, among the many disturbance monitoring methods, the LandTrendr algorithm was designed based on Landsat time-series images, mainly to detect vegetation change trends. Compared with other algorithms, the LandTrendr algorithm is able to detect subtle and low-density changes, and it is the main vegetation change detection method at present [20]. An S-CCD method has been proposed to enhance COLD with a state-space model to improve real-time monitoring of forest disturbances, achieving higher accuracy (F1: 0.793) and 4.4 times faster computation [21]. There are studies using LandTrendr algorithm and Landsat satellite data to describe the forest change patterns in the study area over the years, and the overall identification accuracy of both image change and image stability is very high [22]. By using Landsat satellite data and applying the LandTrendr time-series segmentation algorithm, discriminated the vegetation cover change characteristics and vegetation change patterns in the Yarlung Tsangpo River region while calculating the normalized vegetation index to grasp the vegetation change status of the region from 1984 to 2018 [23].

Forest disturbance is the main factor affecting forest function, structure, growth, and harvest [24] and also affects the stability of the forest ecosystem. The causes of forest disturbance can usually be divided into human factors and natural factors, among which the main influencing factors include forest changes caused by forest fires, deforestation, man-made road construction, forest pests, and diseases [25]. Due to historical reasons and the need for economic development, China’s forests have suffered large-scale destruction. In the 1960s, China’s forest cover area was only 8.7% [26]. Due to the decrease in forest area, the national ecological environment has been greatly affected, and natural disasters such as land desertification, sandstorms, soil erosion, drought, and flood occur frequently, seriously affecting regional ecological security. Forest fires can be triggered by natural factors such as lightning strikes and are often caused by human activities, resulting in the loss of forest structure and biodiversity, while exacerbating soil degradation and carbon emissions, with profound impacts on regional ecosystem services. Deforestation is the direct cause of forest cover loss, closely related to agricultural expansion and timber demand, which leads to habitat destruction, species diversity decline, soil erosion, and other secondary effects. Man-made construction of roads and buildings affects the ecological integrity of forests through habitat fragmentation and induces secondary disturbances such as illegal logging and land encroachment. These three factors have a significant impact on the structure, function, and dynamics of forest ecosystems and are the core challenges of global forest conservation and ecological restoration. With the deepening of people’s understanding of environmental change and the implementation of various policies to return farmland to forest, the communique on China’s land greening status in 2022 issued by the Office of the National Greening Commission shows that by 2022, China’s forest area will reach 231 million hectares, and the forest coverage rate will be 24.02%. In this case, based on the study of vegetation change, this study explored the forest change in the Nanling Corridor over the past 20 years. The detection and attribution analysis of forest vegetation change plays an important role in understanding the trend of forest change in Southern China since the 21st century, realizing the effective utilization and protection of forests, realizing the economic development of forestry, and achieving the goal of “double carbon” under the condition of ensuring the healthy development of forests.

This study helps to explain the domestic and international research status and development trend of forest disturbance monitoring based on remote sensing time series from two aspects of forest disturbance detection and attribution, introduces the change detection method and technology of forest disturbance based on the fusion of remote sensing time-series spatiotemporal spectrum information, and summarizes the attribution method and spatiotemporal spectrum characteristics of forest disturbance based on remote sensing time series. The existing problems in remote sensing monitoring of forest disturbance were analyzed, and some suggestions for future research were given. By analyzing the trend of vegetation change and the causes of vegetation disturbances in the Nanling Corridor, we explored the impacts of human activities and natural factors on the vegetation in the Nanling area. This will provide data support and effective recommendations for the development of future forest management measures in this area. This study covered the following issues:

(1)

Data acquisition and image screening: Landsat de-clouded remote sensing image data from 2001 to 2020 were retrieved through the remote sensing data cloud platform GEE, and images with changes between years were screened to determine the vegetation changes in the Nanling Corridor.

(2)

Disturbance changes assessment and accuracy validation: The disturbance and recovery trend detection (LandTrendr) method and random forest algorithm were utilized to assess the long time-series disturbance changes in vegetation in the Nanling Corridor region, and the classification accuracy and kappa coefficient were validated by visually selecting sample points.

(3)

Trend analysis and management recommendations: This study aimed to analyze the trend of vegetation change and the causes of disturbance in the Nanling Corridor, explore the impacts of human activities and natural factors on the vegetation, and provide data support and scientific basis for the formulation of future forest management measures in the Nanling area.

2. Study Area and Data Sources

2.1. Study Area

The study area (shown in Figure 1) was the Nanling Corridor (108°47′ E–116°38′ E, 29°01′ N–22°37′ N) in Southern China. Nanling is a mountainous region connecting Jiangxi Province and Hunan Province with the two southern provinces (Guangxi and Guangdong). It is connected to the Wuyi Mountains in the east, the Yunnan–Guizhou Plateau in the west, the Luoxiao Mountains in the north, and the JiuLian Mountains in the south [27]. The Nanling area is characterized by the fragmentation of the mountains and the changes in their orientations due to many orogenic movements, which has led to mountains without connecting veins. The basins in the western part of the South Ridge are mostly composed of limestone, forming karst landforms; the basins in the eastern part of the South Ridge are mostly composed of red gravel, forming Danxia landforms through weathering and erosion [28] (Figure 1).

2.2. Research Data

In this study, Landsat-5TM, Landsat-7ETM+, and Landsat-8OLI surface reflectance images were collected for the study period 2001–2020. Images obtained by Landsat-5 are by far the most widely used and effective satellite remote sensing information source related to global Earth resources [29]. The data used in this experiment from 2001 to 2011 were provided by Landsat-5. Landsat-7 carries the Enhanced Thematic Mapper (ETM+) sensor and adds a panchromatic band (PAN band) with a resolution of 15 m to improve the resolution of the infrared band, and the resolution of band 6 is also improved from 120 m to 60 m, so its accuracy is improved compared to Landsat-5. In this study, Landsat-7 provided Earth remote sensing image data from 2012 to 2019 [30]. The Landsat-8 Land Imager (OLI) has nine bands, and compared with Landsat-7, the OLI adds one blue band and one short-wave infrared band. The image resolution and accuracy are significantly improved. The 2020 data for this experiment were provided by Landsat 8 [31]. The above remote sensing data are all collected and processed online through the GEE cloud platform, and radiation correction, geometric correction, and atmospheric correction have been carried out. In addition, the GEE platform was used to assess the quality of the Landsat band for cloud masking.

2.3. Other Auxiliary Datasets

This study also used high-resolution satellite image data from Google and China’s land use data for 30 consecutive years from 1990 to 2021. Among them, the high-resolution satellite image data from Google were mainly acquired via the GEE (Google Earth Engine) platform, which was used to classify the selected sample points and perform attribution analysis [32]. The CLCD (China Land Cover Dataset), which contains China’s land use data for 30 consecutive years from 1990 to 2021, was published by Ms. Xin Huang of Wuhan University. This dataset is based on 335,709 views of Landsat data using the Google Earth Engine [33]. It was used in this study to exclude some samples that were not within its scope.

3. Methods

The aim is to analyze forest disturbance changes and their driving factors in the study area using multi-temporal remote sensing imagery and ground validation. The specific work flow is shown in Figure 2. (1) Study Area Definition and Data Collection: The remote sensing data from the Google Earth Engine (GEE) platform will be utilized to access Landsat cloud-free imagery from 2001 to 2020 [34]. The data will undergo preprocessing, including radiometric calibration, cloud detection, and cloud removal, to ensure the quality of the imagery. (2) Disturbance Detection and Sample Construction: The Landsat disturbance and recovery trend detection algorithm (LandTrendr) will be applied to analyze the multi-temporal remote sensing imagery, extracting the timing, intensity, and recovery trends of forest disturbances in the study area. In addition, visual interpretation will be conducted to validate the disturbance detection results. Sample points will be selected, and a validation dataset will be created through random sampling to assess the accuracy of the LandTrendr results. (3) Analysis of Disturbance Driving Factors: Based on the disturbance detection results, random forest algorithms will be used to analyze the driving factors of forest disturbances. Major factors such as forest fires, deforestation, and man-made road and building construction were selected to analyze forest disturbance because of its destructive effect on forest ecosystems and its high correlation with natural and man-made drivers. Ground survey data, along with remote sensing analysis, will be used to explore the correlation between vegetation changes and disturbance factors in the Nanling region, identifying the main influencing factors. (4) Data Analysis: The spatiotemporal distribution and trends of forest disturbance will be analyzed, assessing differences in disturbance intensity and recovery processes across different regions and time periods. Based on the analysis results, the mechanisms by which human activities and natural factors influence forest vegetation in the Nanling Corridor will be explored, providing scientific support for forest management and decision-making in the study area.

3.1. The LandTrendr Algorithm

LandTrendr (Landsat-based Detection of Trends in Disturbance and Recovery) is a forest disturbance detection algorithm based on Landsat satellite remote sensing time series [35]. It identifies forest disturbance events and subsequent recovery processes by analyzing changes in pixel spectral values over time, particularly anomalies in vegetation indices (such as NDVI) or other spectral features. Utilizing a segmented linear regression approach, the algorithm decomposes the time series into segments representing stable, disturbance, and recovery phases. By analyzing the slopes and change magnitudes of these segments, LandTrendr can accurately pinpoint the timing, intensity, and recovery trends of disturbances [36].

In this study, the vegetation disturbance data of the South Ridge Corridor were extracted using the LandTrendr algorithm in the GEE platform in conjunction with a vector map of the study area boundary. This experiment focused on using the LandTrendr algorithm to extract vegetation disturbance data in the South Ridge Corridor using the normalized burn index (NBR) [37]. The NBR is a remote sensing indicator used to assess the recovery of vegetation or burned ground after a wildfire, as shown in the following equation:

$$\mathrm{N}\mathrm{B}\mathrm{R}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}}$$

(1)

where $\rho$NIR and $\rho$SWIR are the surface reflectance in the near-infrared and short-wave infrared bands of Landsat.

In this study, NBR (normalized burn ratio) index was used for vegetation restoration. Although NBR is traditionally used to assess the impact of fires that burn forests, its application in vegetation restoration research has certain scientific basis. By using the ratio of near-infrared (NIR) and short-wave infrared (SWIR) bands, NBR can effectively distinguish between damaged and undamaged areas after a fire. After the fire, the vegetation in the damaged area decreased, the reflectance of SWIR band was higher, and the reflectance of NIR band was lower, so the NBR value decreased significantly [38]. However, with vegetation recovery, especially in the regeneration stage after fire, NBR value gradually recovers, which is represented by the decrease in SWIR reflectance and the increase in NIR reflectance, thus reflecting the process of vegetation recovery.

3.2. The Random Forest Algorithm

Different sample point characteristics were subsequently attributed to vegetation disturbance patches using a random forest algorithm. The random forest (RF) algorithm was proposed by Breiman et al. in 2001 [39]. It is a machine learning method with an integrated learning supply, and it selects random features for input by randomly applying multiple decision trees to a dataset. The random forest algorithm possesses the advantages of fast training speeds and strong resistance to overfitting and perturbation, and it is easy to make parallelized methods. It is a widely used machine learning method [40]. The random forest algorithm is robust and stable in processing high-dimensional data, complex nonlinear relations, and sample imbalance and its built-in feature importance assessment function can provide a more intuitive ecological explanation for research. While some of the latest machine learning algorithms may show greater accuracy on specific tasks, random forests have clear advantages in terms of computational efficiency, model interpretability, and sensitivity to hyperparameter adjustment, especially for heterogeneous data and limited sample sizes common in ecological research.

The forest disturbance data for each year in the study area were obtained and differentiated via the visual interpretation method, and sample data of the visual interpretation of forest disturbance in the South Ridge Corridor from 2001 to 2020 were collected [41]. Suitable indices and multispectral images were selected for the machine language training of perturbed data to achieve random forest classification. In this study, the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI) were used in random forest classification attribution. The Enhanced Vegetation Index (EVI) is a remote sensing index that reflects the growth status of vegetation by combining the information in the red and near-infrared bands of the vegetation index and eliminating the influence of factors such as vegetation cover, soil background, and the atmosphere on the remote sensing data [42]. Through previous studies on EVI and NDVI, it is demonstrated that NDVI is the most effective in forest disturbance analysis. The NDVI was the most important random forest classification index in this study. The NDVI is an index used to detect the growth status of vegetation and vegetation cover. It can quickly reflect the status of vegetation cover on the ground [43]. In this study, the obtained sample data were used to calculate the NDVI using the following equation:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathsf{\rho }}_{\mathrm{R}\mathrm{e}\mathrm{d}}}{{\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathsf{\rho }}_{\mathrm{R}\mathrm{e}\mathrm{d}}}}$$

(2)

where $\rho$Red and $\rho$NIR are the surface reflectance in the red and near-infrared bands of Landsat. The characteristics of selected sample sites were compared using normalized vegetation index data. This study focused on the attribution of unclassified data by using the NDVI to analyze sample point information. By selecting one of the sample points and calculating the change in the value of its NDVI data during the study period, corresponding to the change in a false-color image of that sample point, it was possible to verify if it was reasonable to use the NDVI data for random forest attribution [44].

Before conducting this analysis, we used the LT-GEE Pixel Time Series to retrieve the NDVI for a specific point from 20 June to 20 September each year between 1984 and 2021, generating a line graph to assess whether the NDVI data accurately reflected vegetation disturbances (Figure 3). The NDVI at this point was relatively flat from 1984 to 1996. It reached a low value in 1997, trended upward from 1998 to 2003, and leveled off from 2004 to 2021 [45]. The obtained NDVI data were explored by comparing them with high-resolution images of the point and its surroundings from Google, and, in 1990, the point and its surroundings were covered with vegetation, so the NDVI was higher that year. The vegetation cover at and near this point decreased in 1997, so the NDVI values were low at this time. In 2003, the vegetation at and near the point recovered but did not reach the level of vegetation cover in 1990. The NDVI was higher in 2003 than in 1997 but not as high as in 1990. These results show that using the NDVI to study vegetation disturbance is justified [46].

In this study, about 600 samples are selected for classifier training every year from 2001 to 2020. A total of 70% of the training points are used for random forest training, and the remaining 30% are used for analyzing classification results. In this study, overall classification accuracy and Kappa coefficient were used to evaluate the classification accuracy. The overall classification accuracy refers to the proportion of the data presented correctly after classification to the total data. Kappa coefficient is an index of correlation between the classified data and the original data based on the confusion matrix. The higher the value of the above index, the higher the classification accuracy. Kappa coefficient equal to 0 indicates that there is no correlation between the two, and the classification accuracy is low. If the value is equal to 1, it means that the classification result is consistent with the verification data [47]. Generally, the Kappa coefficient should be higher than 0.6, indicating that the classification results are highly consistent with the validation data.

4. Results

4.1. Distribution of Spatial and Temporal Patterns of Forest Disturbance in the South Ridge Corridor

4.1.1. Spatial Distribution Patterns of Forest Disturbance in the South Ridge Corridor

There are many mountains in the Nanling Corridor area, so there is a large area of forest distribution. In this study, the vegetation disturbance data of the Nanling Corridor for the past 20 years were extracted from the forest data of the China Land Cover Dataset (CLCD), and a map of the forest disturbance distribution in the Nanling Corridor from 2001 to 2020 was drawn (as shown in Figure 4). The map shows obvious temporal and spatial trends. From the spatial distribution of forest disturbance, it can be observed that there were significant characteristics in the regions where forest disturbance occurred. For example, the frequency of disturbance in the northwestern part of Shaoyang City, the northwestern part of Guilin City, the central parts of Qingyuan City and Shaoguan City, and the southwestern part of Heyuan City was significantly lower than that in other areas, mainly because these areas have many mountains, relatively small populations, low population density, and less forest disturbance caused by human influence. In the southwest of Heyuan City, there is not only a wide range of mountains but also a wide area occupied by Wanlu Lake, which lacks forest distribution, so the forest disturbance was lower. On the other hand, disturbance occurred more frequently at the edges of mountain ranges. In these areas, the forest cover area was relatively high, human settlement areas were relatively close, and the forest was disturbed over a large area to expand the living area and to harvest wood. There were fewer disturbed areas in the middle of Shaoyang City, Yongzhou City, and Shaoguan City, mainly because the vegetation in the non-forest areas was removed and the forest disturbance in the urban areas was lower.

4.1.2. Temporal Distribution Patterns of Forest Disturbance in the South Ridge Corridor

In this study, the forest disturbance in the Nanling Corridor was investigated (as shown in Figure 5). From 2001 to 2020, the area of forest disturbance in the Nanling Corridor was 13,596.92 km², accounting for 5.79% of the total area of the 11 cities in the corridor, and the average annual disturbance area was 679.85 km². From 2001 to 2016, the overall disturbed forest area showed an increasing trend, and the annual disturbed forest area accounted for more than 99% of the disturbed vegetation area, except in 2001, when it accounted for 75.93% of the disturbed forest area. The low forest disturbance area in 2001 may have been due to the serious urban forest disturbance and the high incidence of human-made construction and road construction. The disturbance areas in 2001 and 2008 accounted for 9.53% and 13.40% of the total disturbance area for the past 20 years, respectively. In 2008, the area peaked at 1822.39 km², and 2002 had the smallest area at only 45.48 km², accounting for 0.33% of the total disturbance area for the past 20 years. In 2008, a series of extreme meteorological disasters occurred in the Nanling Corridor, including heavy rains, typhoons, and ice disasters, which caused significant disturbance to the forest ecosystem in the region. Heavy rains and typhoons lead to large-scale soil erosion and tree collapse, which damage the structure and function of forests. Ice disasters further aggravate the physical destruction of forests by overpowering trees with frozen snow. The superimposed effects of these meteorological disasters caused serious damage to forest vegetation, which challenged the ecological restoration of forests and the stability of species habitat. As a result, the forest disturbance data are highly variable. After 2016, the disturbance area showed a downward trend. It was less than 400 km² in 2019, while it rebounded in 2020, but the overall disturbance area showed a weak rebound trend. During the 20-year period, there were seven years in which the area of forest disturbance exceeded 800 km², namely 2001, 2008, 2013, 2014, 2015, 2016, and 2017, indicating that forest disturbance events occurred more frequently in these years or that large-scale forest destruction events occurred, such as fires or freezing damage.

Graphical data show the spatial distribution changes in forest disturbance in three typical regions between 2001 and 2020 (Figure 6). Figure 6a shows historical Google images; Figure 6b,c show the spatial distributions of forest disturbance in 2001 and 2020, respectively. There was a significant spatial difference between the locations of forest disturbance in 2001 and 2020, which reflected the spatial dynamics of forest disturbance in the different time periods and its driving mechanism. Specifically, in 2001, the concentrated areas of forest disturbance were mainly distributed in the vicinity of human settlements but, by 2020, the spatial scope of the disturbance had changed, and the locations of the disturbance had shifted or weakened in some areas. This indicates that the spatial distribution of forest disturbance has had different patterns over the past two decades due to the effects of policy regulation, land use change, and natural environmental factors. The disturbance trends in different regions also provide a scientific basis for further forest management and conservation strategies.

4.2. Analysis of Drivers of Spatial and Temporal Patterns of Forest Disturbance in the South Ridge Corridor

4.2.1. Evaluation Results of Forest Disturbance Attribution Accuracy

Based on the sample data obtained from visual interpretation, LandTrendr and the random forest classification algorithm were used to establish a confusion matrix to attribute forest disturbance data in the Nanling Corridor and to calculate the classification accuracy. According to the confusion matrix of forest disturbance attribution (as shown in

Table 1. Confusion matrix of forest disturbance attribution in Nanling Corridor.

Factor	Cutdown	Fire	Building or Road	Total
Cutdown	1743	44	47	1834
Fire	240	732	36	1008
Anthropogenic	198	56	497	751
Total	2181	832	580	3593

), the Cutdown factor attribution had high accuracy. Among all 1834 sample points defined as felling, only 44 and 47 were classified as fire factors and building or road factors, respectively, in the random forest classification. Among the 1008 sample points of fire factors, 240 were classified as Cutdown factors after random forest classification, mainly because the color contrast of the false-color images was weak in some years, making them difficult to distinguish. Among all 751 building or road factors, 198 were attributed to Cutdown factors, giving the lowest accuracy among all factors, mainly because the characteristics of Cutdown factors are similar to those of building or road factors, while building or road factors tend to last a long time and are easy to attribute to felling factors. This difference in disturbance factor attribution affected the overall accuracy and the Kappa coefficient of forest disturbance data attribution.

The disturbance data of each factor were counted to obtain the overall precision of the forest disturbance attribution and the statistical table of the Kappa coefficient (as shown in

Table 2. Overall accuracy and Kappa coefficient of forest disturbance attribution.

Year	Overall Accuracy	Kappa Coefficient
2001	84.83	0.71
2002	84.47	0.73
2003	79.27	0.66
2004	81.82	0.69
2005	86.23	0.78
2006	83.09	0.71
2007	83.05	0.71
2008	82.47	0.72
2009	76.19	0.61
2010	80.11	0.64
2011	79.33	0.62
2012	84.5	0.68
2013	87.78	0.73
2014	88.44	0.80
2015	83.58	0.72
2016	84.11	0.75
2017	81.54	0.68
2018	81.05	0.67
2019	81.66	0.66
2020	75.98	0.63
Average	82.48	0.70

). Based on the analysis of the attribution data, the total accuracy of the attribution of the forest disturbance data was 82.48% and the Kappa coefficient was 0.70. The overall accuracy for 2003, 2009, 2011, and 2020 was less than 80%, mainly due to incomplete cloud removal and less obvious contrast in the false-color images, resulting in precision deviations during classification. In this study, visual interpretation was used to identify sample types when selecting sample points, and manual interpretation may have a certain degree of deviation for its own reasons related to discrimination. In the selected samples, it was found that the spatial distribution of the selected sample points also affected the classification accuracy. If the density of the selected samples was too high, the classification accuracy was easily reduced. Therefore, in order to maintain classification accuracy, sample points should be selected as evenly as possible in a space.

In this study, the main factors of forest disturbance were divided into three categories, namely deforestation, fire, and buildings or roads. Land use data were used to process forest disturbance data to obtain forest disturbance factor data of the Nanling Corridor from 2001 to 2020 (as shown in

Table 3. The areas and proportions of forest disturbance factors from 2001 to 2020.

Years	Area				Percentage
	Cutdown	Fire	Building or Road	Total	Cutdown	Fire	Building or Road
2001	952.51	262.65	188.69	1403.85	67.85	18.71	13.44
2002	22.52	8.56	6.63	37.71	59.72	22.70	17.58
2003	188.57	55.68	62.72	306.97	61.43	18.14	20.43
2004	212.89	116.03	41.71	370.63	57.44	31.31	11.25
2005	111.05	83.20	29.83	224.08	49.56	37.13	13.31
2006	288.49	96.89	54.14	439.52	65.64	22.04	12.32
2007	280.88	113.55	43.38	437.81	64.16	25.94	9.91
2008	579.44	869.51	104.40	1553.35	37.30	55.98	6.72
2009	275.56	147.64	58.40	481.06	57.22	30.66	12.13
2010	307.57	130.72	66.42	504.71	60.94	25.90	13.16
2011	264.80	150.71	57.16	472.67	56.02	31.88	12.09
2012	345.01	103.11	31.84	479.96	71.88	21.48	6.63
2013	436.01	171.45	89.26	696.72	62.58	24.61	12.81
2014	457.58	246.05	68.94	772.57	59.23	31.85	8.92
2015	428.68	204.92	167.63	801.23	53.50	25.58	20.92
2016	438.44	265.33	152.62	856.39	51.20	30.98	17.82
2017	541.54	172.65	97.73	811.92	66.70	21.26	12.04
2018	395.30	103.96	47.79	547.05	72.26	19.00	8.74
2019	206.24	38.62	68.35	313.21	65.85	12.33	21.82
2020	225.51	74.39	92.47	392.37	57.47	18.96	23.57

Units: km² and %.

). Among the forest disturbance factors, the proportion of deforestation factors decreased, but they still accounted for the highest proportion, reaching 58.16% of all disturbance factors. Compared with forest factors, deforestation factors only decreased slightly in most years. The year with the greatest decrease was 2001, probably because the urban expansion area was larger that year, leading to massive deforestation. The proportion of fire factors increased to 29.06%, which may have been due to their large area after a forest fire. The proportion of building or road factors decreased to 12.78%. When calculating the areas of forest disturbance factors, many urban data points were excluded, resulting in a decrease in the proportion of building or road factors (Figure 7).

In the distribution diagram of the forest disturbance factors (Figure 8), it can be seen that the disturbance areas attributed to building or road factors were relatively concentrated in the whole Nanling Corridor and mostly distributed in clusters or lines. This was because building or road factors mainly occurred during the construction of buildings or roads, and such disturbances tended to occur in urban or expressway areas. The distribution area of fire factors was also concentrated, but it was relatively loose compared with that of building or road factors. This was because fire usually occurred in chunks, but its distribution in the overall forest area was relatively loose and the distribution randomness was high. The Cutdown factors were the most widely distributed in the whole region and were more concentrated in the mountain margin area, mostly in a sheet distribution. Among the forest disturbance factors, the distribution of building or road factors decreased significantly, and their cluster distribution was obvious only in Hezhou, Heyuan, and Qingyuan. In the forest area, the building or road factors mostly had a linear distribution, and they were less distributed in the whole study area, as they mainly had the distribution of built roads. The deforestation factors were distributed in the relatively low-elevation areas around the mountains due to the relatively high human activity in these areas. The fire factors were relatively scattered but mostly distributed in the center of the mountains at higher elevations. It can be seen that the attribution data obtained in this study basically had a higher degree of coincidence in the regions with higher probabilities of occurrence of each factor, which further proved the accuracy of the attribution of forest disturbance in this study.

In order to study the temporal variation in each disturbance factor, dynamic change maps of three typical regions were extracted from 2001 to 2020, and the forest disturbance characteristics and driving factors were further analyzed in detail based on a series of forest protection and restoration projects that occurred around the 2000s.

The above analysis showed that from 2001 to 2020, mangrove disturbance was mainly caused by human characteristics in 95% of the years, and deforestation was always the main component of forest disturbance, but its proportion experienced a significant decline. According to the obtained dynamic distribution diagram of the temporal variation rules of deforestation disturbance factors from 2001 to 2020 (Figure 9), in the early stage, deforestation factors accounted for most of the forest disturbance and were the dominant disturbance factors. In particular, from 2001 to 2008, deforestation accounted for almost 80% of all forest disturbance factors, indicating a central role for deforestation in forest disturbance. The main reasons were the strong demand for timber resources for economic development and the reclamation of forests for agricultural expansion. However, with the in-depth implementation of China’s “natural forest protection project” and “return of farmland to forest project”, commercial logging has gradually been strictly limited, and the natural forest ban policy has also significantly reduced deforestation disturbance. At the same time, the structural adjustment of the forestry industry and an increase in wood imports have further reduced the dependence on domestic forest resources, and the promotion of ecological protection awareness has also played an important role.

According to the temporal distribution of fire disturbance factors from 2001 to 2020 (Figure 10), compared with deforestation factors, the changes in fire disturbance factors were influenced more by natural and human factors, and their disturbance proportion increased from 2006 to 2015, reaching a peak in 2008. This may have been closely related to extreme weather (such as droughts and high temperatures) and man-made activities (such as burning land and tourism). However, with upgrades to the national forest fire warning system and improvements in fighting capacity, as well as patrol work, fire disturbance declined between 2016 and 2020, but extreme weather events still pose a potential threat.

Forest disturbance caused by building or road construction showed a gradually increasing trend (Figure 11). In particular, its proportion increased significantly after 2016. This change was mainly driven by the accelerated urbanization process and the expansion of infrastructure, especially in the economically developed southeastern coastal areas. However, with the implementation of the ecological protection red line policy and the optimization and adjustment of land use, the overall scope of building disturbance has been constrained to a certain extent.

From the perspective of the overall trend, from 2001 to 2020, the dominant role of forest disturbance factors in China gradually shifted from human-driven factors (such as logging and construction) to natural factors (such as fire), reflecting the remarkable effects of national policies, technological advances, and ecological protection measures. However, uncertainties brought about by climate change remain an important challenge for future forest conservation. In particular, extreme weather conditions may increase the risk of fire disturbances. In the future, China needs to continue to strengthen policy regulation, strengthen the prevention and control capacity for forest fires, optimize the coordinated development of urbanization and ecological protection, improve the efficiency of forest resource monitoring and management through scientific and technological means, and lay a solid foundation for achieving the “double carbon goal”. The law of forest disturbance change in this period fully illustrates the comprehensive effects of multiple factors such as policy, economy, and nature and their far-reaching impacts on the ecosystem.

4.2.2. Forest Disturbance Attribution Analysis

The natural factors in the study area are not usually highly variable, but extreme natural events can be very disruptive to the ecosystem. For example, the severe snowstorm of 2008 and the droughts that occurred in 2001, 2003, 2007, and 2019 all caused extensive forest loss [48]. Forest loss due to drought, on the other hand, has a lag and usually shows up within a few years after a drought. The severe 2008 southern snowstorm caused extensive damage to vegetation, and, according to an assessment by the China Forestry Editorial Board, the forest area in Southern China affected by this rain, snow, and ice was greater than 2.27 × 105 km² [49]. As the largest mountain range in Southern China, Nanling also suffered major losses in the ice storm. According to incomplete statistics, during this disaster, the loss of forest ecosystem services in the Nanyang Nature Reserve within the Nanling region amounted to CNY 386 million [50]. In the study area, the area of forest disturbance peaked during the study period in 2008.

According to the National Bureau of Statistics (NBS), which released the 2020 Seventh National Population Census, providing the most significant data on cities above the prefecture level, a total of nine cities in the Nanling Corridor region have resident populations of more than 3 million, with the populations of Shaoyang City and Ganzhou City reaching 6,563,500 and 8,970,000 people, respectively. Five of the eleven municipalities in the Nanling Corridor showed declines in their resident populations in the seventh general population survey compared to the sixth general population survey, namely Huaihua, Shaoyang, Hezhou, Heyuan, and Wuzhou. Population factors have a significant impact on forest change, and an increase in population requiring large amounts of farmland to grow crops requires deforestation. Large urban areas are being built to accommodate more people. The strong correlations between the urbanization rate and forest change indicators are mainly due to the fact that along with the urbanization process, the agricultural population enters the city. The phenomena of land abandonment and returning farmland to forests are common, which reduce forest disturbance [51]. At the same time, however, towns and cities need to build large urban areas to accommodate more people, resulting in an increase or decrease in urban forest disturbance.

5. Discussion

5.1. Forest Disturbance Monitoring

One of the main challenges for LandTrendr in detecting forest disturbances lies at the edges of disturbance events, where the algorithm is prone to under-detection errors. As an annual-scale algorithm, LandTrendr selects only one observation from the summer composite period each year to represent annual spectral values. Designed as a pixel-level algorithm, it treats each pixel as an independent spatial entity and does not utilize spatial information from neighboring pixels to improve detection accuracy. This limitation can result in weaker disturbance signals at the edges of forest disturbance events, increasing the likelihood of missing detections. Additionally, LandTrendr requires parameter calibration for optimal threshold values to ensure accurate performance in different regions [52]. For instance, parameter thresholds calibrated for use in China may not be optimal for other regions. Consequently, when applying the algorithm to new areas, recalibration of the parameters is necessary.

Since the public release of Landsat satellite imagery in 2008, researchers have developed numerous change detection algorithms based on Landsat time-series data. These forest disturbance detection algorithms can generally be categorized into two types: annual-scale and daily-scale methods. Annual-scale algorithms detect the year in which disturbances occur, while daily-scale algorithms can pinpoint the exact dates of disturbance events. Although the improved LandTrendr algorithm can identify disturbance years, it cannot determine specific months or days [53]. To accommodate the frequent winter logging in Southern China, the improved algorithm incorporates winter components. However, for northern Chinese forests, where reference samples may be covered by snow from 1 October to 31 December, both LandTrendr and its improved versions struggle to detect disturbances occurring during the winter. Nevertheless, as only a small portion of northern Chinese forests experienced prolonged snow cover from 1 October, the improved LandTrendr algorithm remains applicable when snow-free and cloud-free observations are available during the synthesis period.

Currently, forest disturbance detection algorithms based on remote sensing time series primarily focus on detecting disturbance events, with few algorithms providing information on disturbance types. Thus, future advancements in forest disturbance detection algorithms using Landsat time-series data should aim not only to improve detection accuracy but also to offer detailed information on disturbance types and other relevant factors.

The LandTrendr algorithm has been widely used in forest ecosystem remote sensing monitoring in recent years. Mao et al. reconstructed the time series of forest disturbance parameters based on the LandTrendr algorithm and combined it with the random forest algorithm for forest height estimation [54], and Katsuto et al. used the LandTrendr temporal segmentation algorithm to extract the key features of land use cover changes and forest disturbances from the annual Landsat time-series data [55]. However, most of these studies focused on analyzing forest disturbances and lacked analysis of the influencing factors that cause forest disturbances, whereas, in this study, we not only applied LandTrendr to extract forest disturbance data for the South Ridge Corridor for the years 2001–2020 but also analyzed the attribution of forest disturbances based on anthropogenic and natural factors. Our study fills the gap in the current literature, provides a multifactor-driven perspective behind forest change, and provides a theoretical basis for the development of forest protection and sustainable development policies.

In addition, the LandTrend algorithm combined with the random forest secondary classification method adopted in this study can only detect forest disturbance events (felling, fire, etc.) at the stand level. For pests and diseases or minor disturbances of less than 900 square meters, the LandTrendr algorithm combined with random forest methods cannot detect them well [56]. Fortunately, a method called COLD has been developed in the research, which is suitable for the detection of different levels of interference in all ground classes, and this method does not require sample points for training.

5.2. Forest Disturbance Driving Analysis

In attributing forest disturbances, the driving factors are highly complex. This study primarily analyzes typical categories such as logging, wildfires, and construction or road development. LandTrendr, a remote sensing algorithm based on time-series data, focuses on detecting the timing, intensity, and recovery of vegetation disturbances by analyzing spectral change patterns. However, the algorithm predominantly relies on spectral information and physical environmental changes, lacking a comprehensive analysis of policy impacts, which limits its ability to fully explain the mechanisms driving forest disturbances.

For instance, in 1979, China’s first environmental resource law, the Forest Law of the People’s Republic of China (for Trial Implementation), was promulgated, followed by the Opinions of the State Council on Further Implementing Policies and Measures on Returning Cultivated Farmland to Forestry (2002), the Decision of the Central Committee of the Communist Party of China and the State Council on Accelerating the Development of Forestry (Zhongfa [2003] (No. 9)), the Outline of the National Plan for the Conservation and Utilization of Forestland (2009), and the 13th Five-Year Plan for Forestry Development (2016). These policies have been effective in reducing deforestation and forest loss in the study area. Since 2000, the central government has introduced a series of policies to liberalize logging quota management. The policies maintain the original logging quota program while providing more flexible quota allocation methods for plantations and commercial forests [57]. As a result of these policies, the area of forest disturbance in the South Ridge Corridor area has been trending upward since 2001. China formally enacted a reform system for the collective forest rights system in 2008, which allowed individuals and enterprises to own and manage forest land, enabling farmers to break through existing logging controls established through power-seeking, poaching, and abusive logging, as well as expanding individuals’ rights to use forest land [58]. Under the auspices of this policy, the area of forest disturbance in the study area increased after 2008 compared to the previous period and then stabilized. By 2015, after the amendment of the Forest Law of the People’s Republic of China, the area of forest disturbance began to show a decreasing trend.

Although LandTrendr effectively detects the spectral characteristics of such disturbances, it cannot associate disturbance signals with specific policy contexts or uncover the intrinsic connections between policy implementation and ecological changes. This limitation hinders the algorithm’s capacity to provide comprehensive insights into the drivers of forest disturbances and restricts its potential application in policy evaluation and forest management. Incorporating policy factors into the disturbance analysis framework, alongside socio-economic data and the spatial distribution of policy implementation, could enhance the algorithm’s capability to reveal the intricate mechanisms driving forest disturbances and offer valuable directions for its improvement.

6. Conclusions

Based on the GEE (Google Earth Engine) cloud processing platform and Landsat data with a 30 m resolution, this study used the LandTrendr time-series change detection algorithm and random forest classification to monitor the spatiotemporal pattern of forest disturbance in eleven cities in four provinces in the Nanling Corridor. The distribution of forest disturbance in the study area was mapped over 20 recent years. During the study period, the forest disturbance area in the Nanling Corridor also had significant characteristics in its temporal pattern. It showed an upward trend from 2001 to 2007, reached its peak in 2008, maintained a steady upward trend from 2009 to 2016, and showed a downward trend thereafter. Based on sample point training via visual interpretation, the classification results of the forest disturbance factors in the Nanling Corridor from 2001 to 2020 were verified. The overall classification accuracy reached 82.48%, and the Kappa coefficient was 0.70, indicating high classification accuracy. Among the forest disturbance factors, the proportion of deforestation factors was highest at 58.45%, and the proportion of construction or deforestation factors was lowest at 12.85%. Among the three types of disturbance factors, the building or road factors were most strongly affected by humans, mainly in the urban areas of the cluster distribution and a small part of the linear distribution. The Cutdown factors were mainly distributed in the edge areas of the mountains at lower altitudes, and their distribution was mostly flaky and dense. The fire factors were mostly concentrated in the center of the mountains with relatively high elevations, and their distribution was relatively loose. This study quantified the regional status and change trend of forest disturbance in the Nanling Corridor. The forest disturbance dataset generated in this study provides detailed information about the temporal and spatial changes in forests in the Nanling Corridor over the past 20 years, addressing gaps left by other studies in the Nanling Corridor over longer periods. The series lasting up to 20 years provide insight into the mechanisms and characteristics of forest change. Timely and accurate monitoring of the forest’s status is conducive to formulating and implementing sustainable management and policies, formulating climate change mitigation policies based on the ecological status, and calculating the carbon balance at the national level [46].