Information Extraction and Prediction of Rocky Desertification Based on Remote Sensing Data

Abstract

Rock desertification has become the third most serious ecological problem in western China after desertification and soil erosion. It is also the primary environmental problem to be solved in the karst region of southwest China. Karst landscapes in China are mainly distributed in southwest China, and the area centered on the Guizhou plateau is the center of karst landscape development in southern China. It has a fragile ecological environment, and natural factors and human activities have influenced the development of stone desertification in the karst areas to different degrees. In this paper, Dafang County, Guizhou Province, was selected as the study area to analyze the effect of the decision tree and multiple linear regression model on stone desertification and to analyze the evolution characteristics of stone desertification in Dafang County from 2005 to 2020. The FLUS model was applied to predict and validate the stone desertification information. The results show that the overall accuracy of multiple linear regression extraction of stone desertification is 70%, and the Kappa coefficient is 0.69; the overall accuracy of decision tree extraction of stone desertification is 60%, and the Kappa coefficient is 0.521. The multiple linear regression stone desertification extraction model is more accurate than the traditional decision tree classification. The overlay analysis of stone desertification and slope, elevation, slope direction and vegetation cover showed that stone desertification was more distributed between 1300–1900 m in elevation; stone desertification decreased gradually with the increase in slope; each grade of stone desertification was mainly distributed in the range of 5 to 25° in slope, which might be related to human activities. The FLUS model was used to predict the accuracy of 2015 data in the region and project the changes in stone desertification area in 2035 under a conventional scenario and an ecological protection scenario in the region to provide a new reference for predicting stone desertification.

1. Introduction

As the degree of rock desertification deteriorates, vegetation cover and biomass decrease. Coupled with the impact of human disturbances, ecological degradation and biodiversity reduction have curtailed social and economic development [1,2,3]. Karst is distributed around the world, widely in densely populated areas such as Eastern Europe, the Middle East, Southeast Asia, and Florida [4]. China accounts for 16% of the global karst area and is one of the countries with the widest distribution of karst landforms [5,6,7]. The distribution of karst in China is concentrated and typical [8]. The most severe area of karst is located in Guizhou province, with the widest distribution of stone desertification and the most significant hazard. Dafang county is known as the “karst kingdom” [9]. It is an ideal study area.

The word “Karst” usually refers to a unique geological type. However, there is some confusion about its meaning and application in relevant foreign literature because the word “Rocky Desertification” is an ancient term from a specific region. According to records [10], “Rocky Desertification” was first proposed at the American Association for advancing science symposium in 1983. Research on rocky desertification at home and abroad mainly focuses on technical research [11,12,13], temporal and spatial evolution of rocky desertification [14,15,16], and research on the rocky desertification microenvironment [17,18,19,20,21,22]. In 2009, Ling Chengxing et al. [23] used the improved EVI index to extract rocky desertification information by analyzing the differences between the spectral features of multispectral data and multiple vegetation indexes and used the NDWI index to modify the water information, combined with spatial slope analysis to carry out decision tree analysis to improve the accuracy of information extraction. The research shows that this method is a fast and effective method for extracting rocky desertification information; In 2011, Zu Qi [24] and others extracted the surface coverage information of the Zhaidi area in Guilin by using the object-oriented classification method; Wang Xiaoxue et al. [25] established the karst model to analyze the contribution rate of human and natural factors. Bao Yan and others used a neural network to predict rocky desertification in a county in Guangxi in combination with terrain, vegetation, humanities, and other factors [26]. The CA-Markov model is mostly used in predicting rocky desertification and is the result of optimizing the data on the CA model. Many scholars have recently optimized the CA model with the neural network, logistic, and Markov algorithms to improve its accuracy. Compared with the traditional simulation, the FLUS model integrates neural networks and Markov to improve classification accuracy [27]. The FLUS model optimizes the defects of the CA-Markov model, is more objective, and is widely used in land-use changes. This study applies the model to rocky desertification prediction, explores the feasibility of the FLUS model to predict rocky desertification, and provides a new idea for rocky desertification prediction.

The management of stone desertification is the key to achieving sustainable socio-economic development and the critical factor in liberating and promoting regional economic development. Scientific understanding of the spatial and temporal evolution characteristics of stone desertification is the primary premise and foundation to start the management of stone desertification. Therefore, based on remote sensing technology to extract information on stone desertification, we studied the spatial distribution characteristics of stone desertification and the development mechanism of stone desertification in Dafang County to provide the theoretical basis for the management of stone desertification areas. Secondly, the evolution characteristics of stone desertification are studied to test the effectiveness of stone desertification management in the study area. Finally, the FLUS model is used to predict the quantity and spatial distribution characteristics of stone desertification in 2035 to indicate the management direction of stone desertification and to provide a new reference for predicting stone desertification.

2. Materials and Methods

2.1. Overview of the Study Area

Dafang County is located in the middle of Bijie City, Guizhou Province, in the middle of the Yunnan–Guizhou Plateau, as shown in Figure 1. 105°15′47″–106°08′04″ E; 26°50′02″–27°36′04″ N, with an elevation range of 1400–1900 m. The distance between east and west is 862 km, and the distance between north and south is 852 km [28]. The county’s total area is 2746 square kilometers, accounting for 13.04% of the total area of Bijie City and 1.99% of the total area of Guizhou Province, which is an inland mountainous county near the sea. Dafang County is a mountainous agricultural county with more prominent karst terrain, and each township has different degrees of stone desertification land distribution, which seriously restricts economic and social development. Dafang County has a subtropical humid monsoon climate, with an average annual temperature of 11.8 °C, a maximum temperature of 32.7 °C and a minimum temperature of –9.3 °C. The average annual precipitation is 1155 mm, with precipitation mainly concentrated from April to September, accounting for about 78.8%. Two rivers are distributed in the county, namely the Chishui River and the Wujiang River, with a total length of 727.3 km, mainly in the study area’s southern, central and western parts.

2.2. Experimental Design and Data Source

The Landsat-5 data 2005, 2015 and 2019, and 2020 Landsat-8 data were selected. The data series were systematically radiometrically corrected and geometrically corrected right after downloading, so only radiometric calibration and atmospheric correction were required. The slope data were extracted by 30 m resolution DEM data, which requires three-dimensional imagery for complete coverage, and the DEM data were mosaicked and cropped. Combining the actual elevation range of 700–2300 m in the study area, the elevation was classified into six classes < 1, 1–1.3, 1.3–1.6, 1.6–1.9, 1.9–2.2, 2.2–3 km with reference to the spatial and temporal distribution of stone desertification in the karst region of southeastern Yunnan by Ma Li-Chi et al. [29]. Combined with the slope range of 3–74° in Dafang County, the slopes were classified into 6 classes flat, gentle, sloping, steep, sloping, sharp, and dangerous, i.e., 0–5°, 6–15°, 15–25°, 26–35°, 36–45°, and ≥46°. Land use type data were obtained from the 2020 land use type data of the Institute of Space and Sky Information Innovation, Chinese Academy of Sciences (https://data.casearth.cn/ (accessed on 6 February 2022)). Precipitation data from 2005–2020 were used to calculate elevation information, obtained through the Resource and Environment Science and Data Center, Chinese Academy of Sciences (https://www.resdc.cn/ (accessed on 6 February 2022)) to download. Finally, the resolution and spatial reference of each element were unified.

2.3. Research Methods

2.3.1. Decision Tree Construction

Based on the research results of Zuo Tai’an and others on rocky desertification in the Bijie area, combined with the regional characteristics of Dafang County, this study takes the rock exposure rate and vegetation coverage as the main indicators to divide the degree of rocky desertification and establishes the classification standard of rocky desertification intensity [30].

(1)

Vegetation coverage (FVC)

Vegetation cover (FVC) is the percentage of the vertical projection of the area of all vegetation in the growing area to the area of the study area [31]. The lower the vegetation cover, the more severe the soil erosion, the larger the exposed area of bedrock, and the more severe the degree of rock desertification; the higher the vegetation cover, the less severe the soil erosion, and the smaller the exposed area of bedrock, and the less severe the degree of rock desertification.

$$\mathrm{NDVI}=\frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{Red}}$$

(1)

where NDVI is an index reflecting the status of land cover vegetation, NIR is the near-infrared band, and Red is the red band.

$$\mathrm{FVC}=\frac{\mathrm{NDVI}-{\mathrm{NDVI}}_{\mathrm{R}}}{{\mathrm{NDVI}}_{\mathrm{R}}-{\mathrm{NDVI}}_0}$$

(2)

where ${\mathrm{NDVI}}_0$ is the NDVI value with a cumulative frequency of 5%; ${\mathrm{NDVI}}_{\mathrm{R}}$ is the NDVI value with a cumulative frequency of 95%.

The vegetation cover degree of the whole area is divided into three parts; when $\mathrm{NDVI}<{\mathrm{NDVI}}_{\mathrm{O}}$ and $\mathrm{NDVI}>{\mathrm{NDVI}}_{\mathrm{O}}$, FVC takes 0 and 1, respectively, and when ${\mathrm{NDVI}}_{\mathrm{O}}>\mathrm{NDVI} >{\mathrm{NDVI}}_{\mathrm{R}}$, FVC is calculated using the following equation:

$$\mathrm{FVC}=\left({\mathrm{b}}_1{ \mathrm{lt} \mathrm{NDVI}}_0\right) \times 0 + \left({\mathrm{b}}_1{ \mathrm{gt} \mathrm{NDVI}}_{\mathrm{R}}\right) \times 1 + \left({\mathrm{b}}_1{ \mathrm{ge} \mathrm{NDVI}}_0{ \mathrm{and} \mathrm{b}}_1{ \mathrm{le} \mathrm{NDVI}}_{\mathrm{r}}\right) \times [\frac{{\mathrm{b}}_1-{\mathrm{NDVI}}_0}{{\mathrm{NDVI}}_{\mathrm{r}}-{\mathrm{NDVI}}_0}]$$

(3)

where ${\mathrm{b}}_1$ is the NDVI. The formulae for the years 2005, 2015, 2019, and 2020 are shown in

Table 1. FVC Calculation Formula of Each Year.

Year	FVC Calculation Formula
2005	$\left({\mathrm{b}}_1 \mathrm{lt} 0.18\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} 0.52\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} 0.18{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} 0.52\right){ \times \left[\right(\mathrm{b}}_1-0.18)∕(0.52–0.18)$]
2015	$\left({\mathrm{b}}_1 \mathrm{lt} 0.22\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} 0.76\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} 0.22{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} 0.76\right){ \times \left[\right(\mathrm{b}}_1-0.22)∕(0.76–0.22)$]
2019	$\left({\mathrm{b}}_1 \mathrm{lt} 0.22\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} 0.68\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} 0.22{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} 0.68\right){ \times \left[\right(\mathrm{b}}_1-0.22)∕(0.68–0.22)$]
2020	$\left({\mathrm{b}}_1 \mathrm{lt} 0.29\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} 0.87\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} 0.29{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} 0.87\right){ \times \left[\right(\mathrm{b}}_1-0.29)∕(0.87–0.29)$]

below.

(2)

Rock exposure rate (FR)

The rock exposure rate (FR) is the percentage of exposed rocks in the study area, excluding the area covered by the vertical projection of vegetation [9]. It is also due to the large area of rock exposure on the surface that the karst area has formed a complex combination of small habitat types and diverse combinations, with relatively little land cover, forming a poor and arid living environment, and with human interference, degenerating into a desert-like landscape of karstic rock desertification. Rock exposure is the final stone desertification landscape phenomenon. In studying the stone desertification evolution process, bedrock exposure is the most direct influence factor and the most intuitive natural phenomenon to judge the change of stone desertification phenomenon.

$$\mathrm{NDRI}=\frac{\mathrm{SWIR}-\mathrm{NIR}}{\mathrm{SWIR}+\mathrm{NIR}}$$

(4)

The NDRI is an index reflecting the exposed condition of land rocks, SWIR is the short-wave infrared band, and NIR is the near-infrared band.

According to the principle of the image dichotomy model, the NDRI of each time phase in the study area is firstly calculated by using the above formula. Then, the statistical data of NDRI images are calculated. Assuming a confidence level of 1, the rock exposure rate (“FR”) is calculated as follows:

$$\mathrm{FR}=\frac{\mathrm{NDRI}-{\mathrm{NDRI}}_{\mathrm{R}}}{{\mathrm{NDRI}}_{\mathrm{R}}-{\mathrm{NDRI}}_0}$$

(5)

where NDRI₀ is the NDRI value with a cumulative frequency of 5%; NDRI_R is the NDRI value with a cumulative frequency of 95%.

When $\mathrm{NDRI}<{\mathrm{NDRI}}_0$ and $\mathrm{NDRI}>{\mathrm{NDRI}}_0$, FR is 0 and 1, respectively; when ${\mathrm{NDRI}}_0>\mathrm{NDRI}>{ \mathrm{NDRI}}_{\mathrm{R}}$, FR is calculated using the following formula:

$$\mathrm{FR}=\left({\mathrm{b}}_1{ \mathrm{lt} \mathrm{NDRI}}_0\right) \times 0+\left({\mathrm{b}}_1{ \mathrm{gt} \mathrm{NDRI}}_{\mathrm{R}}\right) \times 1+\left({\mathrm{b}}_1{ \mathrm{ge} \mathrm{NDRI}}_0{ \mathrm{and} \mathrm{b}}_1{ \mathrm{le} \mathrm{NDRI}}_{\mathrm{r}}\right) \times [\frac{{\mathrm{b}}_1-{\mathrm{NDRI}}_0}{\left({\mathrm{NDRI}}_{\mathrm{r}}-{\mathrm{NDRI}}_0\right)}]$$

(6)

where b₁ is the NDRI image. The formula for calculating FR for each year is shown in

Table 2. FR Calculation Formula of Each Year.

Year	FR Calculation Formula
2005	$\left({\mathrm{b}}_1 \mathrm{lt} -0.47\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} 0.04\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} -0.47{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} 0.04\right) \times [\frac{{\mathrm{b}}_1+0.47}{\left(0.04+0.47\right)}]$
2015	$\left({\mathrm{b}}_1 \mathrm{lt} -0.58\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} –0.04\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} -0.58{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} –0.04\right) \times [\frac{{\mathrm{b}}_1+0.58}{\left(–0.04+0.58\right)}]$
2019	$\left({\mathrm{b}}_1 \mathrm{lt} -0.48\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} –0.02\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} -0.48{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} –0.02\right) \times [\frac{{\mathrm{b}}_1+0.48}{\left(–0.02+0.48\right)}]$
2020	$\left({\mathrm{b}}_1 \mathrm{lt} -0.69\right) \times 0 + \left({\mathrm{b}}_1 \mathrm{gt} –0.15\right) \times 1 + \left({\mathrm{b}}_1 \mathrm{ge} -0.69{ \mathrm{and} \mathrm{b}}_1 \mathrm{le} –0.15\right) \times [\frac{{\mathrm{b}}_1+0.69}{\left(–0.15+0.69\right)}]$

below.

Taking into account the availability of data and the characteristics of the study area, and combining the regional characteristics of Dafang County, the rock exposure rate and vegetation coverage are taken as the main indicators for the division of rocky desertification degree, and the grading standard of rocky desertification intensity is established, as shown in

Table 3. Grading Standards for Rock Desertification Intensity.

Strength Grade	Rock Exposure Rate (%)	Vegetation Coverage (%)	Utilization Value
Nothing	0–40	70–100	Construction land, water body, forest land, grassland, etc.
Potential	40–60	50–70	Forest land, shrubland, grassland, etc.
Light	60–70	35–50	Sparse shrubbery, sloping farmland, grassland, unused land, etc
Moderate	70–80	20–35	Sparse shrubbery, sloping farmland, grassland, unused land, etc.
Severe	80–90	10–20	Stone-sloping farmland, grassland, unused land, etc.
Extremely severe	90–100	0–10	Stone-sloping farmland, grassland, unused land, etc.

.

2.3.2. Construction of Multiple Linear Regression Model

Using a multiple linear regression model to extract stone desertification means extracting waveform information and calculating NDVI and NDRI after selecting six types of training areas, namely, heavy stone desertification, moderate stone desertification, potential stone desertification, light stone desertification, water bodies, vegetation and buildings, and testing the correlation between NDVI, NDRI and topographic factors and stone desertification classification, as shown in

Table 4. Correlation test.

Correlation Factor	Correlation Coefficient	Correlation Factor	Correlation Coefficient
NDVI	−0.820	DEM	−0.475
NDRI	−0.308	Aspect	0.057
Slope	−0.327	Rocky desertification grade	1

. In this study, slope, slope direction, elevation, NDVI, and NDRI were selected as influencing factors for the correlation test; the strongest correlation was NDVI, and the correlation coefficient reached 0.820; NDRI, slope, elevation and classification grade were weakly correlated, all were about 0.3, while slope direction and the classification correlation coefficient was the smallest at 0.057 not correlated, so NDVI, NDRI, Slope, and DEM were selected as independent variables affecting the distribution of rock desertification.

The rocky desertification information extraction model is obtained by the multiple linear regression of NDVI, NDRI, slope and DEM (

Table 5. Remote sensing information extraction model of rocky desertification.

Image	Formula for Classification of Rocky Desertification Types	R2
Landsat5	$7.968-7.943\times \mathrm{NDVI}+0.057\times \mathrm{NDRI}-0.023\times \mathrm{Slope}-0.002\times \mathrm{DEM}$	0.822
Landsat8	$6.459-4.88\times \mathrm{NDVI}+0.094\times \mathrm{NDRI}-0.021\times \mathrm{Slope}-0.001\times \mathrm{DEM}$	0.701

):

$$\mathrm{RDL}=\mathrm{a}+\mathrm{b}\times \mathrm{NDVI}+\mathrm{c}\times \mathrm{NDRI}+\mathrm{d}\times \mathrm{Slope}+\mathrm{e}\times \mathrm{DEM}$$

(7)

where a, b, c, d and e are constants.

In this study, the Landsat-8 remote sensing images are combined in pseudo color according to the short wave infrared (Band 6), near-infrared (Band 5) and red light band (Band 4). The Landsat-5 remote sensing images are combined in pseudo color according to the mid-infrared (Band 5), infrared (Band 4) and red light band (Band 3) to distinguish buildings, vegetation, non-vegetation and water bodies. The study area is divided into non-rocky desertification, potential rocky desertification, moderate rocky desertification and severe rocky desertification, including water, buildings and forest land.

2.3.3. FLUS Model

The FLUS model is based on social and natural impacts to predict land change and the future. The principle of this model is derived from cellular automata (CA). The ANN model and adaptive inertial competition model are added based on the traditional Markov model. Firstly, the FLUS model uses a neural network algorithm to obtain the suitability probability of each land use type within the research scope from the first phase of land use data and multiple driving factors, including human activities and natural effects; Secondly, the FLUS model can better avoid error transmission by sampling the land use distribution data of Phase I. In addition, in the process of land change simulation, the FLUS model mechanism can effectively deal with the uncertainty and complexity of the mutual transformation of various land-use types under the common influence of natural actions and human activities, which gives the FLUS model higher simulation accuracy and can obtain results similar to the real land use distribution.

(1)

The calculation formula for the suitability probability of the ANN neural network model is:

$$S\mathrm{P}\left(\mathrm{p},\mathrm{k},\mathrm{t}\right)={{\displaystyle \sum }}_{\mathrm{j}}\mathrm{wj},\mathrm{k}\times \mathrm{sigmoid}\left({\mathrm{net}}_{\mathrm{j}}\left(\mathrm{p},\mathrm{t}\right)\right)$$

(8)

$$={{\displaystyle \sum }}_{\mathrm{j}}\mathrm{wj},\mathrm{k}\times {\left[1+{\mathrm{e}}^{-\mathrm{netj}\left(\mathrm{p},\mathrm{t}\right)}\right]}^{-1}$$

(9)

where $\mathrm{SP}\left(\mathrm{p},\mathrm{k},\mathrm{t}\right)$ represents the suitability probability of land type K at spatial position P when t, WJ, and K are the adaptive weights of the hidden layer and the output layer; $\mathrm{sigmoid} ({\mathrm{net}}_{\mathrm{j}}\left(\mathrm{p},\mathrm{t}\right)$ is the excitation function from the hidden layer to the output layer. $\mathrm{netj}\left(\mathrm{p},\mathrm{t}\right)$ is the feedback signal of the neurons in the jth hidden layer on the training practice t and pixel p. For the suitability probability output by the neural network model, the sum of the suitability probabilities of various land-use types is constant as 1. This provides the probability distribution data of land suitability in 2015 and 2020.

(2)

The basic formula of the Markov model is:

$${\mathrm{S}}_{\left(\mathrm{t}+1\right)}={\mathrm{S}}_{\left(\mathrm{t}\right)}\times {\mathrm{P}}_{\mathrm{ij}}$$

(10)

where ${\mathrm{S}}_{\left(\mathrm{t} +1\right)}$ represents the number of predicted patches at t + 1; ${\mathrm{S}}_{\left(\mathrm{t}\right)}$ represents the number of predicted patches at t; ${\mathrm{P}}_{\mathrm{ij}}$ represents the transition probability matrix.

Markov is used to predict the quantity for 2035. For this prediction, the rocky desertification data in 2015 and 2020 was imported to predict 2035, and the number of pixels of various types of rocky desertification in 2025, 2030 and 2035 were obtained.

(3)

Adaptive inertia competition model formula:

$${\mathsf{\Omega }}_{\mathrm{p},\mathrm{k}}^1=\frac{{{\displaystyle \sum }}_{\mathrm{n}\times \mathrm{n}}\mathrm{con}\left({\mathrm{c}}_{\mathrm{p}}^{\mathrm{t}-1}=\mathrm{k}\right)}{\mathrm{N}\times \mathrm{N}-1}\times {\mathrm{w}}_{\mathrm{k}}$$

(11)

where, ${\mathsf{\Omega }}_{\mathrm{p},\mathrm{k}}^1$ represents the domain image of land type K on pixel P at the t iteration; ${\displaystyle \sum }_{\mathrm{n}\ast \mathrm{n}}\mathrm{con}\left({\mathrm{c}}_{\mathrm{p}}^{\mathrm{t}-1}=\mathrm{k}\right)$ The total number of pixels of land type K in the domain window of N; ${\mathrm{w}}_{\mathrm{k}}$ is the domain weight of different land use types.

The distribution of rocky desertification in 2015 was predicted in 2005, the distribution data of rocky desertification in 2005 and the probability distribution data of land suitability in 2015 were imported, and the total number of pixels of each rocky desertification type in 2015 was inputted. In this study, the number of iterations was set as 100, and the neighborhood type is 3 × 3. After the operation was completed, random accuracy verification was carried out to judge the accuracy of the prediction results. Similarly, the probability distribution data of land suitability in 2020 is imported, and the number of each type of rocky desertification pixel in 2035 is input. The transfer probability matrix was set to 1, and the neighborhood weight was set to 1. After the operation, the conventional scenarios of rocky desertification in 2035 were obtained.

3. Results

3.1. Spatial and Temporal Evolution Characteristics of Rocky Desertification in Dafang County

The decision tree and multiple linear regression model are established for rocky desertification to extract rocky desertification information, which is applied to the images of four phases. Finally, the spatial–temporal distribution of rocky desertification in Dafang County from 2005 to 2020 is shown in Figure 2.

During 2005–2020, the area of stone desertification decreased, and the area of stone desertification decreased from 997.16 km² in 2005 to 320.63 km² in 2020, with a total decrease of 676.53 km². The area of potential stone desertification, moderate stone desertification and severe stone desertification all decreased, and the area of no stone desertification increased, and the potential, moderate and severe stone desertification evolved towards no stone desertification during 2005–2020.

Comparing the classification results with the actual types, quantifying the accuracy of the classification results, and accuracy testing are important prerequisites to ensure the reliability of the data. To quantitatively analyze the accuracy of decision tree classification and multiple linear regression model in extracting rocky desertification information, the overall classification accuracy and kappa coefficient were used to test the accuracy of the 2019 rocky desertification classification map. Samples were randomly collected for each rocky desertification level and visually interpreted using high-definition satellite images. A total of 126 samples were collected by decision tree classification in this study, including 24, 22, 21, 20, 19 and 20 random sample points of extremely severe, severe, moderate, mild, potential and those free from rocky desertification, respectively.

The spatial distribution of sample points is shown in Figure 3. The spatial distribution of 29, 30, 37 and 33 random sample points of severe, moderate, potential and rocky desertification free in the multiple linear regression model is shown in Figure 3. The overall classification accuracy of decision tree classification of rocky desertification is 60%, and the kappa coefficient is 0.521. The overall classification accuracy of the multiple linear regression model for rocky desertification is 70%, and the kappa coefficient is 0.69. The multiple linear regression model has higher accuracy in extracting rocky desertification than the decision tree classification. The multiple linear regression model is more suitable for extracting stone desertification data in Dafang County than the decision tree.

Given the field situation combined with the distribution map and sample point map, it can be seen there was no continuous distribution of rock desertification vegetation or obvious bedrock exposed; the potential rock desertification vegetation cover was high with bare land slightly exposed; there was moderate rock desertification to mainly scrub grass bushes mainly, and the vegetation cover had obvious low bare land. The heavy rock desertification vegetation was sparse and contained mainly scrub grass with the patchy distribution.

To analyze the characteristic map made inverse by the multiple linear regression model, the area and percentage of stone desertification in Dafang County in 2005, 2015, 2019, and 2020 are shown in the following table. Therein, it can be seen that the area without stone desertification in 2005 is the widest, accounting for 45% of the national land area of Dafang County, the area of stone desertification (area of moderate stone desertification + area of severe stone desertification) accounts for 36% of the national land area of Dafang County, of which the area of severe stone desertification The area of heavy stone desertification is 413.09 km², accounting for 15%, the potential stone desertification is 18%, and the moderate stone desertification is 21.27%. According to Figure 4, it can be seen that in 2005, the heavy stone desertification in Dafang County was continuously distributed in strips, and it was commonly distributed in Dafang County.

In 2015, the area without stone desertification was 1827.766 km², accounting for 67%. Stone desertification accounted for 22% of the total area of the study area, of which the area of heavy stone desertification was 296.36 km², accounting for 11%, the potential stone desertification accounted for 12%. The medium stone desertification accounted for 11.03%. Figure 4 shows that in 2015, the heavy stone desertification in Dafang County was distributed in small patches, mainly concentrated in the western and southern areas, such as Lihua Township and Yangchang Township.

The area without stone desertification in 2019 was 1712.72 km², accounting for 62%. Stone desertification accounts for 21% of the county area, of which 8% was heavy stone desertification, 17% was potential stone desertification area, and 12.73% was moderate stone desertification. According to Figure 4, it can be seen that in 2019, the heavy stone desertification in Dafang County was distributed in small patches, mainly in the southern, central and western areas, such as Lihua Township, Huang Nantang Township, Cat Farm Township, Walnut Township, Ringshui Township, etc.

In 2020, the area without stone desertification was 2231.15 km², 81%, and stone desertification accounts for 12% of the county area, whose heavy stone desertification accounted for 5%, potential stone desertification accounted for 7%, and moderate stone desertification accounted for 7.08%. Figure 4 shows that in 2020, the heavy stone desertification in Dafang County was distributed in small patches, mainly in the southern area, i.e., Lihua Township, Huang Nantang Township, Machang Township and other areas.

Combined with Figure 2 and

Table 6. Stone desertification grade area and percentage.

Year	2005		2015		2019		2020
Grade of Rocky Desertification	Proportion (%)	Area (km2)	Proportion (%)	Area (km2)	Proportion (%)	Area (km2)	Proportion (%)	Area (km2)
Nothing	45.32	1244.42	66.56	1827.77	62.37	1712.72	81.25	2231.15
Potential	18.37	504.42	11.62	319.11	16.97	465.99	7.07	194.22
Moderate	21.27	584.06	11.03	302.77	12.73	349.54	7.08	194.55
Severe	15.04	413.09	10.79	296.36	7.93	217.76	4.59	126.07
Total	1	2746	1	2746	1	2746	1	2746

, the area of stone desertification showed a decreasing trend from 2005 to 2020. The area of potential stone desertification, moderate stone desertification and heavy stone desertification all decreased, and the area of no stone desertification increased. The potential, moderate and heavy stone desertification evolved toward no stone desertification from 2005 to 2020.

The formation and development of stone desertification in the karst area are the results of natural and human-made effects, and the distribution characteristics of stone desertification are revealed by combining the actual situation of Dafang County. The slope, elevation, slope direction, vegetation cover and stone desertification distribution map are overlaid and analyzed to reveal the distribution characteristics of stone desertification and the formation mechanism of stone desertification in the study area by using the decision tree method, extracting slope and slope direction data from DEM data and calculating vegetation cover by remote sensing images, as shown in Figure 4.

The analysis shows that stone desertification in Dafang County is mainly distributed at 1.3–1.6 m above sea level, and the area of stone desertification is more distributed in this elevation range, followed by the range of 1.6–1.9 m; while the distribution of stone desertification is less in the elevation level <1.3 and >1.9 km. The rivers are distributed in areas with lower elevations, and the areas with higher elevations are often mountain tops with less human activities and larger vegetation cover, so the stone desertification area is less. This may be related to human activities. In the area with a gentle slope, it is less difficult for a human to reclaim the land and unreasonable human activities make the stone desertification area higher; on the contrary, the steeper the slope, the more difficult it is to develop and use and thus less human activity, and most of the area has been implemented to return the land to forest. The probability of stone desertification is low.

The relationship between the distribution of stone desertification and the slope is not uniform and can be divided into two categories. One group of scholars believes that the grade of stone desertification increases with the increase in slope; the other group believes that when the slope reaches a threshold value (more often believed to be 25°) the phenomenon of stone desertification will be alleviated. The slope direction indirectly influences the factors of stone desertification under the effect of solar radiation and precipitation, and moderate and severe stone desertification in the east, southeast, south and southwest directions are widely distributed in karst areas. However, the difference in human activity intensity in each direction is not great, so the difference in stone desertification area in each broken direction is not significant. The increase in vegetation cover can improve the current stone desertification to a certain extent. On the contrary, the increase in the stone desertification area will aggravate the deterioration of the stone desertification phenomenon to a certain extent. Therefore, the increase in vegetation cover in the densely vegetated areas inhibits the deterioration of the stone desertification phenomenon.

3.2. FLUS Model Prediction

China’s development goals may be challenging for 15 years from 2020 to 2035 to realize socialist modernization based on building a moderately prosperous society in all aspects, and the report points out to promote green development and harmonious coexistence between humans and nature in 2035. Therefore, this study uses the evolution characteristics of stone desertification from 2015 to 2020 as the basis for spatial prediction and quantitative prediction for 2035 and also uses the data from 2005 to predict the distribution of stone desertification in 2015 thus for accuracy check.

Various parameters and basic rocky desertification distribution data are imported into the FLUS model to predict the accuracy of the actual data and prediction data in 2015. The kappa value is 0.71, and the overall accuracy is 0.69. The simulation effect is good. Considering that the rocky desertification control policy launched by Dafang County will have a significant impact on the evolution of rocky desertification in the future, two cases of conventional development and rocky desertification control policy are set in the FLUS model, to compare and analyze the similarities and differences of rocky desertification development under ecological protection. According to the change law of rocky desertification from 2005 to 2015, the distribution of rocky desertification in Dafang County under the two scenarios of conventional scenario and ecological protection is adopted, and the statistical differences are shown in Figure 5.

In the absence of human disturbance, the distribution of rocky desertification in 2035 is predicted based on the data in 2020. Under the conventional scenario, the FLUS prediction results show that there will be certain changes in all levels of rocky desertification in Dafang County in 2020 and 2035. The area without rocky desertification will increase significantly, and the area of severe, moderate and potential rocky desertification will decrease. The proportion of rocky desertification will be reduced from 12% today to 11%.

Due to a series of rocky desertification control measures issued by the government of Dafang County, the transition probability matrix under the conventional scenario is adjusted, and the distribution of rocky desertification in the study area under ecological protection in 2035 is predicted by the FLUS model. The change law of rocky desertification at all levels is consistent with the conventional scenario. Still, under the condition of ecological protection, the change rate of rocky desertification area at all levels is larger, which the rocky desertification area reduced from 12% today to 8%; therefore, to achieve more effective control results, we should adhere to the implementation of rocky desertification control policies and continue to uphold the concept that green water and green mountains are golden mountains and silver mountains.

Table 7. Comparison of prediction results of rocky desertification in 2035 under two scenarios.

Type	Rock Free Desertification		Potential Rocky Desertification		Moderate Rocky Desertification		Severe Rocky Desertification
	Area (km2)	Proportion (%)	Area (km2)	Proportion (%)	Area (km2)	Proportion (%)	Area (km2)	Proportion (%)
General scenario	2304.31	0.84	147.16	0.05	168.64	0.06	125.88	0.05
Ecological protection scenario	2399.06	0.87	119.24	0.04	141.02	0.05	86.68	0.03
Variation	94.75	—	−27.92	—	−27.62	—	−39.20	—

shows the statistical table of rocky desertification areas and proportion under the conventional and ecological protection scenarios. It can be seen that the rocky desertification-free area under the ecological protection scenario is higher than that under the traditional scenario, with a change of 94.75 km². The area of potential rocky desertification, moderate and severe rocky desertification, is lower than the conventional scenario, with an average change of 31.58 km².The rocky desertification area was reduced from 11% under the conventional scenario to 8% under the ecological protection scenario. The study found that this difference is mainly distributed in the areas with higher altitudes, so it is predicted that the Rocky Desertification in the areas with higher altitudes will be further improved under the ecological protection scenario. It is speculated that the reason for this phenomenon is that there are fewer human activities in the areas with higher altitudes, and the vegetation is less likely to be disturbed under the ecological protection scenario, resulting in the decline of rock exposure rate and the reduction of rocky desertification area.

4. Discussion

Analyzing the previous research on the evaluation and classification of rock desertification as the basis, a decision tree was established to extract rock desertification using vegetation cover and rock exposure rate, which is limited by the empirical knowledge of experts and requires a rich knowledge reserve. Decision trees, as a kind of binary tree model classification algorithm based on knowledge discovery and data mining, are applied in many fields by establishing a hierarchical level of tree structures to classify one by one layer by layer; the accuracy of this classification method is determined by the rationality of the classification logic hierarchy organization, consisting of root nodes, intermediate nodes and terminal nodes, the decision tree recurs the data set to small branches after testing at branch nodes, and finally outputs the results. Therefore, a multiple linear regression model was constructed to extract information on rock desertification. A total of six training areas of heavy rock desertification, moderate rock desertification, potential rock desertification and vegetation, buildings, and water bodies are selected to explore the correlation between slope, slope direction, elevation, normalized vegetation index, normalized rock index and classification results, and a multiple linear regression model is constructed. The high-definition satellite images were checked for accuracy, and the optimal rock desertification extraction method was selected. The classification results were superimposed with slope, slope direction, elevation and vegetation cover to explore the distribution characteristics and formation mechanism of rock desertification; the principle of constructing an information extraction model to extract rock desertification information is simpler operable and more accurate.

(1)

The research shows that the kappa coefficient of rocky desertification extracted by the decision tree is 0.521, and the overall classification accuracy is 60%. The kappa coefficient of rocky desertification extracted by the multiple linear regression model is 0.69, and the overall classification accuracy is 70%. Therefore, the multiple linear regression model extracts stone desertification data with higher accuracy compared with the decision tree.

(2)

The distribution of rocky desertification is analyzed by superposition with terrain and vegetation factors. The results show that the rocky desertification of each grade has specific distribution characteristics under each slope, elevation, aspect and vegetation coverage. The slope is inversely proportional to rocky desertification, and the vegetation coverage is inversely proportional to rocky desertification. The Rocky Desertification in Dafang County is mainly distributed at an altitude of 1300 m~1900 m.

(3)

The FLUS model predicts that the area without rocky desertification will increase in 2035, and the area of severe, moderate and potential rocky desertification will decrease. Under the conventional scenario, the rocky desertification area is reduced from 12% to 11%. The area of rocky desertification under the ecological protection scenario is reduced from 12% to 8%, and the change rate of rocky desertification under the environmental protection scenario is greater.

The rocky desertification area is decreasing year by year. The changing trend of the rocky desertification area is in line with the annual control area of rocky desertification issued by the Dafang County Government, and significant control results have been achieved.

5. Conclusions

This study extracts stone desertification information by comparing decision tree classification with a multiple linear regression model, establishes a stone desertification classification system, analyzes the spatial and temporal evolution pattern of stone desertification in the past 15 years, and predicts the distribution of stone desertification in 2035. This study points out the direction for stone desertification management and ecological restoration. There are some shortcomings in the study, which can be further studied and improved.

(1)

Due to the low availability of Landsat data from 2000–2020 in Dafang County, only four-time phases of stone desertification monitoring data were collected in 2005, 2015, 2019 and 2020, with a time interval of 15 years, and additional periods are needed at a later stage with time to explore the succession pattern of stone desertification.

(2)

The FLUS model prediction only collects some driving factors and is closely related to human subjectivity and policy implementation; thus, more influencing factors can be considered in the follow-up to improve prediction accuracy.