1. Introduction
Karst rocky desertification is a major global ecological and environmental problem, threatening the sustainable development of human society [
1]. It is a process of land degradation under the background of fragile karst geology and a humid or semihumid climate in tropical and subtropical regions [
2,
3] and is a special type of land desertification. Under the joint actions of the natural environment and human activities [
4,
5,
6,
7,
8], soil and water losses occur, land productivity decreases, and soil erosion is severe, resulting in large exposed areas of basement rock, showing landscape characteristics similar to desertification [
5,
6,
7].
The karst area in Southwest China is located in the center of the karst area in Eastern Asia, which is one of the most widely distributed and developed karst landforms in the world [
8,
9,
10,
11]. Karst rocky desertification is the primary ecological and environmental problem in the karst mountainous area of Southwest China, which restricts the sustainable development of the social economy. It is one of the main causes of poverty in Southwest China and seriously threatens the living environment of local residents [
4,
12]. Guizhou is located in the southwest continuous karst concentration area and is becoming a main karst research area in China. The monitoring and control of rocky desertification and the determination of its impact on the ecological environment in Guizhou are the hot topics in current research [
13,
14].
The development of remote sensing technology has enabled the large-scale monitoring of vegetation cover and rocky desertification, which has led to advances in ecological environment monitoring in karst areas [
3,
15,
16,
17,
18]. Improvements in the precision of automatic mapping with remote sensing technology are important for finding the feature factors that can effectively express karst rocky desertification. Albedo is an effective rock apparent factor. Wei et al. [
19] used Landsat 8 to construct a feature space model of albedo and NDVI, MSAVI, and TGSI and compared their extraction accuracies. The results showed that the overall classification accuracy of the three models was more than 80%, which indicated that the three feature space models are feasible for extracting desertification information from the Mongolian Plateau. The normalized difference rock index (NDRI) is also a suitable factor for expressing karst rocky desertification. For example, the NDRI was used to improve the bedrock exposure index of a system to determine the types and spatial distribution of karst rocky desertification [
20]. Mokhtar et al. [
21] found a close relationship between evapotranspiration index and NDVI and the evapotranspiration index can reflect the growth of vegetation. Deng et al. [
22] retrieved the land surface temperature (LST) of a karst mountain area using Landsat 8 and discussed the relationship between LST and the LUCC and NDVI of different land use types. The results showed that there were significant differences in the LST of different land use types, with the highest being recorded for construction land and the lowest being recorded for forests. The spatial distribution of NDVI and LST showed an opposite pattern. Zhang et al. [
23] found that KRD were primarily associated with the land use, gravel content, and slope gradient in a typical karst plateau (Houzhai River Basin). Li et al. [
24] found that there is a positive correlation coefficient between the KRD index and slope in the karst area, especially where the slope degree is greater than 18°, the trend toward strong rocky desertification is apparent. In this study, after referring to the relevant literature, we conducted a correlation analysis between the rocky desertification grade and the factor set of the samples to find the most effective characteristic factors to express karst rocky desertification.
Supervised classification is one of the effective methods used to extract rocky desertification from remote sensing images. Mixed pixel decomposition is one of the key technologies of supervised classification. The binary classification pixel is a common method used to solve mixed pixel decomposition, which is stable and scientific. Su et al. [
25] studied the relationship between the NDVI value and vegetation coverage in Guangxi using a binary pixel model with a TM image and extracted and classified rocky desertification information based on decision tree theory. The overall classification was more than 85%. You et al. [
26] proposed a local adaptive multiendmember spectral hybrid analysis algorithm to extract subpixel rocky desertification information from medium-resolution images. The results showed that it is accurate and reliable in estimating rocky desertification information. Zhao et al. [
27] established the calculation model of vegetation coverage in Guangxi with MODIS EVI data using a pixel dichotomy model and analyzed the spatiotemporal dynamics process and characteristics of rocky desertification from 2004 to 2014. Zhang et al. [
28] used Landsat-8 Operational Land Imager (OLI) images to construct a novel karst rocky desertification index (KRDI) as an indicator of rocky desertification based on greenness, humidity, and brightness. They found that combining spectral information with spatiotemporal information is a promising method for extracting information on karst rocky desertification. Qi et al. [
29] used ALOS images, the dichotomous pixel model (DPM), and spectral hybrid analysis (SMA) methods to improve the accuracy of rocky desertification estimation in Southwest China up to 80.5%.
In summary, several problems remain in terms of the data sources and special environment of natural geography: (1) The above methods are more limited to small and medium-sized areas such as rocky desertification pilot areas and county areas, and there are few studies on large-scale macro provinces. (2) Due to the difficulty of sample collection, there are not enough samples for modeling or model validation. In most studies, samples have usually been collected and validation was performed by visual interpretation. (3) Most scholars used the analytic hierarchy process (AHP) and the underground detector method for weight analysis and accumulation to enable quantitative remote sensing monitoring of rocky desertification by extracting the vegetation coverage, bedrock bare rate, elevation, land use, and other related data. The model constructions have been simple and the robustness or accuracy of the models has been insufficient.
In this study, we presented an improved method to give accurate mapping of the rocky desertification in Guzhou province, China by considering the vegetation type and seasonal characteristic of the vegetation index. Multiresource data, including surface reflectance, land surface temperature (LST), normalized vegetation index and enhanced vegetation index (NDVI/EVI), transpiration land cover types, and so on, were used as the driven variables, while China’s National Forest continuous inventory data were used as the ground truth to predict rocky desertification. Three data-driven models, including the logistic regression model, random forest (RF) model, and support vector machine (SVM) model, were developed and tested. Furthermore, the temporal and spatial patterns of rocky desertification in Guizhou province were analyzed, which provided a scientific reference for the making of ecological management planning of the study area [
30,
31].
5. Discussion
Rocky desertification monitoring can provide reliable data support for precise rocky desertification control and ecological restoration. The accuracy of rocky desertification monitoring depends on the accuracy of the model and reliable samples and an efficient model algorithm to help improve the quality of the rocky desertification monitoring. In most previous studies, models were constructed using factors such as FVC, RE, and slope, with weight analysis and cumulative calculation, which were then classified; these models produced low rocky desertification monitoring and classification accuracy. In this study, different vegetation heights and vegetation seasonal phases were fused to build a nonlinear machine learning model based on sufficient ground truth and multisource remote sensing data, so the accuracy of our proposed model was higher.
5.1. Effects of Different Vegetation Types on Rocky Desertification Monitoring
The difference in vegetation heights was considered in the optimization of our rocky desertification estimation model. As a result, the accuracy of the estimation model was significantly increased, from 80.6% to 86.4%, and the Kappa coefficient was improved from 0.707 to 0.793. This shows that the different vertical heights of vegetation had a close relationship with rocky desertification in this study. The average vegetation heights of different grades of rocky desertification areas (especially the difference between the non-karst and the karst rocky desertification areas) were obviously different. The soil conditions were good and the possibility of rocky desertification was relatively low in areas with high vegetation growth. Vegetation rarely grows into tall and dense trees in karst rocky desertification areas; even if some vegetation cover is present, it is mainly composed of small shrubs or grass. However, the difference in the spectral characteristics between low shrubs and medium and high trees in the images was low; that is, the different ground objects had the same spectrum, which affects the accuracy of the rocky desertification estimation model.
In this study, land cover type was divided into three categories in the model. The first type included the categories of farmland, buildings, and water, which were removed before modeling because they did not occur in rocky desertification areas. However, if not removed, the buildings and terraces in high-altitude areas are often misjudged as rocky desertification areas, which affects the accuracy of the model. The second type included dwarf shrub and grassland, i.e., low-level vegetation. Due to the poor soil and water conditions in rocky desertification areas, shrubs, mosses, and herbaceous vegetation easily grow on the rock surfaces or crevices. The third type included tall arbor forest land, i.e., high-level vegetation. These areas have fertile soil and good water and soil conditions, so rocky desertification does not easily occur. This vegetation type division modeling can avoid rocky desertification misjudgment, i.e., the problem of different objects having the same spectrum.
Although there were enough samples in the original sample data, there were not enough sample points of extremely severe rocky desertification. Therefore, the severe rocky desertification and extremely severe rocky desertification were combined in the rocky desertification model. In future research, we will attempt the accurate discrimination of severe and extremely severe rocky desertification using the sample equilibrium method. Limited by the unbalanced number of original samples, the vegetation types were divided into two levels, high and low, according to the difference in the vertical height of vegetation growth in this study. This classification method is rough and does not consider the more precise division of vertical height of grassland, shrub, coniferous forest, broad-leaved forest, and other vegetation. More samples will be obtained to achieve more precise differentiation of vegetation types in future research to improve the accuracy of rocky desertification estimation.
5.2. Effects of Seasonal and Temporal Differences on Rocky Desertification Monitoring
In rocky desertification areas, vegetation growth was worse, the vegetation index was low, and the difference in different season phases was minimal, especially in moderate and severe rocky desertification areas. The soil nutrients were better in the non-rocky desertification area, the vegetation growth conditions were better, and, generally, there were medium and tall trees. Therefore, the vegetation index of these sample areas was higher and the differences in the seasonal phases in deciduous forest areas were obvious. The accuracy of the rocky desertification monitoring model was effectively improved by incorporating the seasonal differences in the vegetation index. The standard deviation of the vegetation index in the four seasons was calculated and used as the basis for further classification modeling in this study, which improved the accuracy of the rocky desertification monitoring model from 86.4% to 91.1%, and the Kappa coefficient was improved from 0.793 to 0.861. Xu et al. [
63] used SVM to classify rocky desertification areas and the results showed that the overall accuracy in Liujiang county was 85.50% and the Kappa coefficient was 0.81; the overall accuracy in Changshun county was 84.0% and the Kappa coefficient was 0.79; the overall accuracy in Zhenyuan County was 84.46% and the Kappa coefficient was 0.81.
Since satellite remote sensing data are vulnerable to cloud interference, it is often necessary to synthesize multiphase data to obtain the complete satellite data of a study area. Therefore, we divided it into four seasons, spring, summer, autumn, and winter, to calculate the seasonal temporal vegetation index, and each season had six periods of data synthesis to ensure that the data were basically free from cloud interference. However, there may be some errors in using only four groups of data (spring, summer, autumn, and winter) to determine the change in the vegetation index over a year. In the future, the data filling method for areas with cloud interference will be researched and more groups of data will be used to express the temporal differences of samples in a year in order to improve the accuracy of temporal data and to improve the accuracy of the rocky desertification monitoring model.
In addition, we only used the seasonal temporal data as the basis of classification modeling and errors may have occurred in distinguishing evergreen forest land and rocky desertification areas. The seasonal temporal differences and the vegetation index characteristics of the samples will be combined as the basis of classification modeling to optimize the estimation model of rocky desertification in future research.
5.3. Mixed Pixels and Scale Effect
Scale effect is an important factor that affects the rocky desertification monitoring model. We tried to reduce the error caused by the scale effect in the following three aspects:
5.3.1. Ground Survey Data
NFCI data were used as the ground-measured data of the rocky desertification monitoring model in this study. The size of the sample survey area was 25.8m × 25.8 m, but the recorded rocky desertification grade data were not only based on the size of survey area; we conducted a comprehensive evaluation of the rocky desertification around the survey area. Therefore, the ground survey data were reliable in this study.
5.3.2. MODIS Data
Affected by the lithology, structure, and human activities, the distribution of rocky desertification had obvious regionality and most of them were contiguous. The average area of rocky desertification patches is 0.45 km
2; the resolution is about 212 m [
64,
65]. The original spatial resolution of the MODIS satellite remote sensing data used in this paper ranged from 250 to 1000 m. In the research process, the resolution was unified to 250 m using the resampling method, which was basically consistent with the spatial distribution characteristics of rocky desertification.
5.3.3. Google Earth Engine
Google Earth high-definition image data were used for visual interpretation assistance and each grade of the rocky desertification samples was randomly selected for visual interpretation in this study. We selected rocky desertification in Guizhou as the research object and used MODIS images as research data. This met our research needs under the 250 m resolution standard. Based on the above three aspects, the error of the scale effect of the data used in this study was low and the research results were reliable. Instrument stability, light saturation imaging, and shadows in optical sensors may also have caused problems, so SAR data can be integrated to compensate for the shortcomings of optical images in future research. The fusion of high-resolution remote sensing images (such as Landsat) and MODIS images not only retained the advantages of the temporal resolution of MODIS data but also had the advantages of medium to high spatial resolution, which met the accuracy requirements of different-scaled research better.
5.4. Spatial Distribution Characteristics of Rocky Desertification
The MODIS and ground survey data of 2010 were used for modeling in this study. After accuracy verification, the model was extended to other years and we obtained results for a 2001–2019 rocky desertification time series. The model was able to accurately extract the spatiotemporal variation pattern in rocky desertification, which will help the government to implement projects such as precise rocky desertification control and ecological restoration.
Due to the complex interactions between the different levels of rocky desertification, each area presents changes in its characteristics, so the evaluation of the overall changes in rocky desertification cannot be judged by the change in a single rocky desertification level area but should synthesize the change characteristics of each level of rocky desertification. The total area of rocky desertification is decreasing and the overall level is generally weakening due to natural restoration and the implementation of various ecological projects, especially the rocky desertification special project. However, steep slope reclamation and overgrazing are still occurring in a few areas, which are creating new rocky desertification areas. The degree of rocky desertification is different in different areas in Guizhou. The areas with severe rocky desertification are mainly distributed in the west and south of Guizhou, while the degree of rocky desertification is lower in the east and north. This is closely related to the geographical environment and the intensity of local rocky desertification control. The degree of rocky desertification continued developing in a good direction in Guizhou from 2001 to 2019: the proportion of rocky desertification in all levels showed a downward trend and the area of non-rocky desertification area continued to increase.
The spatial distribution characteristics of rocky desertification in Guizhou are obvious, showing a spatial pattern of being heavy in the west, light in the east, heavy in the south, and light in the north. This is consistent with the conclusion of previous studies [
39,
63]. Each grade of rocky desertification is mainly transforming to non-rocky desertification and potential rocky desertification, and the transformation area of mild, moderate, and severe rocky desertification is relatively small, which is consistent with the results of Zhao and Bai [
27,
39,
66]. In 2010, the area of severe rocky desertification was 4228.008 km
2, accounting for 2.4% of the whole province; the area of moderate rocky desertification was 10,570.02 km
2, accounting for 6% of the whole province; the area of slight rocky desertification was 21,140.04 km
2, accounting for 6% of the whole province; the potential rocky desertification area was 45,803.42 km
2, accounting for 26% of the whole province; the area of non-rocky desertification was 94,425.512 km
2, accounting for 53.6% of the whole province. These results are similar to those of Wang et al. [
67]. The model results in 2005 and 2015 were compared with the NFCI data in the same year, and the overall accuracy was more than 90%, indicating that the proposed rocky desertification monitoring model was reliable. Although the visual interpretation was conducted in 2001 and 2019, there may be some deviations due to the lack of ground survey data. However, the accuracy of MODIS remote sensing data quantitative processing was high and this difference might be small, which needs further verification in the future.
6. Conclusions
The MODIS data were used to extract vegetation coverage and bare bedrock rate. Vegetation seasonal phase data and vertical height difference data were fused in this study. Based on the NFCI data, rocky desertification classification models were built using a logical linear model, a random forest model, and a support vector machine model. Mapping of rocky desertification using multisource remote sensing data were conducted in Guizhou. By introducing vegetation types and seasonal changes, the results have shown that the optimized SVM rocky desertification monitoring model was the most accurate, with an overall classification accuracy of 91.1%, kappa coefficient of 0.861, AD value of 0.060, and QD of 0.029. The degree of rocky desertification in Guizhou was significantly reduced and the total proportion areas of LRD, MRD, and SRD have decreased from 24.86% to 15.2% from 2001 to 2019.
In this study, MODIS data was used to generate the rocky desertification level map at 250 m resolution, which was consistent with the continuous distribution scale of rocky desertification in nature, though there may still be mixed pixels [
68]. In the future, more attention should be paid on the mixture pixel problem by using the linear spectral unmixing method to further improve the performance of the rocky desertification mapping. With the rapid development of high resolution remote sensing, a high resolution optical and SAR image can be integrated in the future research, and more state-of-the-art methods, such as deep learning algorithm, can be used to further improve the spatial resolution and accuracy of the rocky desertification mapping [
69,
70]. In addition, the future study will explore the temporal and spatial pattern change of rocky desertification from the aspects of transferring area and transferring rate [
15]. The driving-force analysis of the rocky desertification level will also be conducted, so as to provide scientific decision support for the making of ecological protection planning [
69,
70,
71,
72].