EN中文

Journal of Resources and Ecology >

2023 , Vol. 14 >Issue 4: 833 - 846

DOI: https://doi.org/10.5814/j.issn.1674-764x.2023.04.016

Revegetation and Management of Mines

Remote Sensing Estimation Methods for Determining FVC in Northwest Desert Arid Low Disturbance Areas based on GF-2 Imagery

  • XUE Xinyue , 1 ,
  • GUO Xiaoping , 1, * ,
  • XUE Dongming 1 ,
  • MA Yuan 2 ,
  • YANG Fan 1
Expand
  • 1. School of Soil and Water Conservation, Beijing Forestry University, Beijing 100083, China
  • 2. Seabuckthorn Development and Management Centre of the Ministry of Water Resources (Soil and Water Conservation Plant Development and Management Centre of the Ministry of Water Resources), Beijing 100038, China
*GUO Xiaoping, E-mail:

XUE Xinyue, E-mail:

Received date: 2022-08-20

  Accepted date: 2023-03-02

  Online published: 2023-07-14

Supported by

Key Research and Development Program of China(2017YFC0504406)

Fold

Abstract

Fractional vegetation cover (FVC) is a vital indicator of surface vegetation. Studies of regional vegetation cover are helpful for understanding the status of the regional ecological environment and can provide important references for the formulation of ecological restoration plans and the evaluation of restoration effects. In vegetation cover related research, studies on extraction methods have attracted much attention. Studies have shown that the universality of vegetation cover extraction methods is poor, as well as the existing studies were mostly conducted on agricultural and forest land in wet, semi-humid and semi-arid areas, while few have investigated arid areas with sparse vegetation that is mainly shrubs and grass. To investigate the accuracy and applicability of different methods for estimating vegetation cover in the near-natural zone of the northwest arid desert, this study extracted six vegetation indices (NDVI, SAVI, MSAVI, ARVI, EVI, and MVI), which could effectively exclude soil and meteorological information to obtain pure vegetation information based on GF-2 multispectral-panchromatic fusion images. Two types of models were then established, including the single VI models (DP model) and multi-VI models (R model, RF model and PCA model), three statistics (SSE, R2, RMSE) were introduced to validate model accuracy and four-fold cross-validation was used to probe the models for overfitting. After filtering the models through these methods, the selected model was applied to invert the vegetation coverage in the study area. The results show three key aspects of this system. (1) Among the various models, the DP model constructed using the EVI and the RF model are more suitable for FVC extraction in the study area. This conclusion was further verified by the significant correlation between the inversion results of the FVC for the entire study area by applying these two models. (2) The values of pure bare soil and vegetation pixels (VIs and VIv) in the DP model will obviously affect the accuracy of the model. Thus, the empirical values should not be blindly adopted in actual research. (3) The vegetation distributions in the figures of the FVC results are similar to the outline of the mountains in the study area, indicating that the coverage distribution may be greatly affected by topographic factors. It is recommended that this aspect should be introduced in subsequent studies.

Key words： FVC; GF-2; random forest model

Cite this article

XUE Xinyue , GUO Xiaoping , XUE Dongming , MA Yuan , YANG Fan . Remote Sensing Estimation Methods for Determining FVC in Northwest Desert Arid Low Disturbance Areas based on GF-2 Imagery[J]. Journal of Resources and Ecology, 2023 , 14(4) : 833 -846 . DOI: 10.5814/j.issn.1674-764x.2023.04.016

1 Introduction

Fractional vegetation cover (FVC) is not only a vital indicator of surface vegetation condition, but also an important parameter for characterizing the ecological environment and climate change, The FVC value is usually defined as the percentage of vegetation area occupying a sample area after vertical projection of the ground. The study of FVC can help us to visualize the trends of changes in regional habitats and provide a crucial reference for the development of regional ecological restoration plans and the evaluation of restoration effectiveness.
The research on FVC covers a wide range of topics, among which the research on FVC extraction methods has attracted the attention of many experts and scholars. The commonly used means of FVC estimation can be classified into two types: ground measurements and remote sensing methods. Field surveys, including sample square methods, visual estimation methods, instrumental measurement and photographic methods, are labor-intensive and inefficient during the whole process. In contrast, remote sensing technology, which mainly includes UAV and satellite remote sensing (Liu et al., 2021a; Wang et al., 2021), has been widely applied to the dynamic study of FVC due to its advantages of high monitoring efficiency, large coverage area and limited disturbance (Zhao et al., 2018). In general, there are several methods for extracting FVC by remote sensing, such as multiple linear regression, mixed pixel decomposition, machine learning method, and others (Feng et al., 2017). The accuracy of FVC extraction methods can be largely constrained by the study region, and the universality of these methods is poor (Lv, 2018). Therefore, it is crucial to conduct research on the screening of extraction methods for different regions. The existing vegetation coverage extraction studies are extensive, and the extraction objects are mainly natural forests (Tan et al., 2022), urban forests (Zhou et al., 2021; Wang et al., 2022a) and crop vegetation (Gu et al., 2020; Meng et al., 2020). However, most of the selected research areas are distributed in humid (Liao et al., 2022; Zhang and Feng, 2022) and semi-humid (Zhang, 2022) areas, while there are few relevant studies on arid natural or low human disturbance vegetation areas.
The arid desert area in northwest China is the main area of coal resources in China, so it has become the key vegetation restoration area of ecological restoration projects. Probing the vegetation status in this area, especially in the areas with less human disturbance around the mine sites, will provide an important reference for the overall ecological construction in this region (Wang, 2021a). However, due to the sparse and loose vegetation caused by the low perennial precipitation, the vegetation spectral information in the arid area is easily confused with that of the bare ground surface, which leads to large estimation errors and great estimation difficulty (Ye, 2017). In recent years, many researchers have explored suitable methods for remote vegetation estimation in arid areas (Gao et al., 2018; Shen et al., 2019; Cui et al., 2021; Deng et al., 2021). With the deepening of research, a variety of vegetation indices (Li et al., 2015; Wang, 2021b) and estimation models (Ji, 2018; Chen et al., 2019; Pi et al., 2021) have been proposed, however, the applicability of multiple method models in this study area still needs to be determined.
Remote sensing is still the mainstream method for FVC extraction, and the selection of remote sensing data sources will largely affect the extraction results and accuracy. At present, the data acquisition sources mainly include UAV, satellites and LiDAR. UAVs acquire data with high spatial and temporal resolution as well as high targeting, but they are also limited by regional topography, network limitations and other conditions. LiDAR is an actively emerging remote sensing means that can reflect the structural characteristics of vegetation, and it has some applications in FVC extraction (Su et al., 2017; Tian et al., 2018), but its use is costly and it suffers from the limitations of point cloud data, i.e., the data generated needs to be combined with other remote sensing means (Jia et al., 2013). In contrast, satellite remote sensing data are open and can be quickly acquired at a large spatial and temporal scale, which makes them widely used in FVC research and this source still has a good application prospect. Commonly used satellite data include the Landsat series (Su et al., 2018), MODIS series, SPOT series (Gao and Xu, 2017; Ren et al., 2017), Sentinel series (Wu et al., 2021; Wang et al., 2022b), Worldview (Chen and Lei, 2018), “Resources” series (Zheng et al., 2022), “Gaofen-1” (Zhao et al., 2020), “Gaofen-6” (Chen and Lei, 2018), and other satellites. The “Gaofen-2” satellite is the first sub-meter high resolution civil satellite developed by China. It has red and near-infrared bands that reflect most of the vegetation information, so it can be used for FVC extraction studies. In fact, existing studies using this satellite image source are still relatively scarce, so its applicability still needs to be further explored.
In this study, the low disturbance area near the Wuhai Mining Area in Inner Mongolia is taken as the study area, and the vegetation index (VI) from GF-2 remote sensing images is extracted and combined with ground field survey data obtained in 2021. Using this VI data, the applicability of the single exponential pixel binary model and multiexponential regression models in study area are compared, the most suitable method for the near-natural area of northwest arid desert region is identified, and the results provide a reference basis for the planting of restoration vegetation under reasonable FVC in the mine area.

2 Data and methods

2.1 Study area

The low disturbance area on the east side of the mining area in Wuhai, Inner Mongilia was selected as the study area (Fig. 1). It is located at the junction of Wuhai and Ordos, and belongs to the territory of Ordos City, with a geographic coordinate range of 106°53°28E-107°02°42E, 39°34°08N- 39°47°46N. The total area of the study area is about 212 km2, and the average altitude is about 1621 m.
The study area is located in the upper section of the Yellow River. It belongs to the temperate continental arid climate, and has a large temperature difference between morning and evening. The annual sunshine hours in Wuhai can reach 3138.6 h, and the average annual rainfall is less than 160 mm while the evaporation can reach 3500 mm. In addition, the district resides at the junction of Mu Us Sand, Ulan buh and Hobq Desert, with frequent windy and dusty weather which has an annual average wind speed of 3.1-4.7 m s-1.
The main soil types in study area are windy sandy soil, gray desert soil and brown calcium soil, and the soil exhibits severe salinization and vegetation scrubbing, with poor soil water retention. The zonal vegetation types are mainly desert vegetation type, arid steppe vegetation type, sandy vegetation type and steppe desert vegetation type (Yue et al., 2011). The main plant species include shrub species such as Tetraena mongolica, Ammopiptanthus Mongolicus, Sarcozygium xanthoxylon, Reaumuria songarica, and herbaceous species such as Artemisia desertorum, Halogeton glomeratus, Suaeda glauca, Achnatherum splendens, among others (Liu et al., 2021b).

2.2 Data sources and processing

2.2.1 Ground sample data

In this study, ground field surveys and measurement experiments were conducted from July 23 to August 9, 2020. The sampling time was basically synchronized with the acquisition time of the remote sensing images. The study area was accessed from the Xinxing Coal Mine in Inner Mongolia. The proposed area was found to have little human disturbance so it was suitable to serve as the study area. At the same time, according to the characteristics of sparse and uniform vegetation distribution and the irregular shape of the slope in the study area, sample points were randomly selected and laid out in accordance with the five-point sampling method (Fig. 2). A total of 78 sample points were set (Fig. 3), 77 of which were valid sample points, yielding a total of 385 small sample squares (each 1 m by 1 m in size). The information recorded included the latitude, longitude, elevation, and slope direction of the sample points; the data information of vegetation type, plant species, and plant sub-coverage; and for the total coverage of each small sample square, the average cover of five small sample squares was considered as the cover of the sample point.
Fig. 2 Five-point sampling method
Fig. 3 Distribution of sample points for the actual measurements

2.2.2 Remote sensing data sources and processing

The remote sensing images used in this study were obtained from the GF-2 satellite data acquired on August 7, 2021. The spatial resolution of the satellite reaches 0.8 m in the panchromatic band, 4 m in the multispectral band, and up to 1 m after fusion. It is the highest resolution civil land observation satellite in China. The images were screened through the Natural Resources Satellite Remote Sensing Cloud Service Platform (http://sasclouds.com/chinese/normal/) to find the vegetation growing season (June-September 2021) images that were required for this study. Three scenes of the images (Table 1) were obtained through an agreement with the Chinese Academy of Forestry.
Table 1 Image Information
No. Satellite Sensor Product level Acquisition date Longitude of scenic center (E) Latitude of scenic center (N)
1 GF2 PMS2 LEVEL1A 2021-08-07 107° 39.84°
2 GF2 PMS2 LEVEL1A 2021-08-07 106.94° 39.67°
3 GF2 PMS2 LEVEL1A 2021-08-07 106.88° 39.49°
The image preprocessing can be divided into four steps: 1) Image cropping; 2) Radiometric calibration, atmospheric correction and orthorectification for multispectral images; 3) Radiometric calibration and orthorectification for panchromatic images; and 4) The geometric registration and image fusion process between multispectral and panchromatic images. The resolution of the image after fusing is significantly improved.
GF-2 images contain four bands: red, green, blue and near-infrared. Combining the characteristics of the bands with the sparse vegetation in the study area, seven vegetation indexes (VIs) were selected and calculated using the Band math tool of ENVI Classic 5.3 software (Table 2): Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Modified Soil Adjusted Vegetation Index (MSAVI), Atmospherically Resistant Vegetation Index (ARVI), Enhanced Vegetation Index (EVI) and Modified Vegetation Index (MVI) (Luo et al., 2005; Zhao et al., 2021).
Table 2 Calculation formulas of the vegetation indexes (VIs)
Vegetation index Calculation method Formula serial
number		 Index range Citation Features
NDVI NDVI=$\frac{NIR-R}{NIR+R}$ (2) [-1, 1] Rouse et al.,
1973		 Ability to eliminate most variations in irradiance related to instrument calibration, sun angle, topography, cloud shadows and atmospheric conditions (most widely used)
SAVI SAVI=$\frac{NIR-R}{NIR+R+L}\left( 1+L \right)$ (3) [-1, 1] Huete,
1988		 The soil brightness index L was introduced to create a simple model that could properly describe the soil-vegetation system. Huete (1988) suggested that L be taken as 0.5
MSAVI MSAVI=$\frac{2NIR+1-\sqrt{{{\left( 2NIR+1 \right)}^{2}}-8\left( NIR-R \right)}}{2}$ (4) [-1, 1] Qi et al.,
1994		 Reduces the effect of bare soil in SAVI, and its L value can be automatically adjusted with the vegetation density, better at eliminating the soil background effect
ARVI ARVI=$\frac{NIR-2R+B}{NIR+2R-B}$ (5) [-1, 1] Kaufman and Tanre, 1992 Reduces sensitivity to atmospheric effects by normalizing radiation in the blue, red and near-infrared bands and corrects remote sensing data for molecular scattering and ozone absorption
EVI EVI=$\frac{2.5\left( NIR-R \right)}{NIR+6R-7.5B+1}$ (6) [-1, 1] Bannari
et al., 1995		 Enhances the vegetation signal by adding a blue band to correct the effect of soil background and aerosol scattering, more suitable for dense vegetation areas
MVI MVI=$\sqrt{\frac{NIR-R}{NIR+R+0.5}}$ (7) [0, 1] McDaniel and Haas, 1982 Eliminates or weakens the effect of soil background on vegetation reflectance

Note: NIR, R and B are the reflectances of the features in the near-infrared, red and blue bands, respectively. L is the soil brightness index between 0 and 1, usually taken as 0.5.

2.3 Method

2.3.1 Single VI model

Dimidiate pixel model (DP model): The pixel is the base unit used by satellite sensors to record the spectral signal of the ground surface. In general, a pixel often contains information about multiple features with different spectral characteristics, each of which contributes to the digital number (DN) value (Ding, 2015). The dimidiate pixel model (DP model) only considers dividing each pixel into two components, vegetation and soil, and it then uses this idea to calculate vegetation coverage, so it is simple and widely used. Gutman and Ignatov (1998) combined the NDVI index with the DP model by viewing the NDVI value of a pixel as the linear sum of the NDVI values of vegetation and soil, and obtained formula (1).
$NDVI=f\times NDV{{I}_{v}}+\left( 1-f \right)NDV{{I}_{s}}$
where $NDV{{I}_{s}}$ represents the NDVI value of the pure bare soil pixels; $NDV{{I}_{v}}$ represents the NDVI value of the pure vegetation pixels; and $f$ is FVC, which is obtained by transforming formula (1) into formula (8).
$f=\frac{NDVI-NDV{{I}_{s}}}{NDV{{I}_{v}}-NDV{{I}_{s}}}$
The NDVI values of pure bare soil and pure vegetation will directly affect the accuracy of the model (Li, 2003). Many studies have defined the NDVIv value as the DN at a cumulative contribution of 95%, and the NDVIs as the DN at 5%, based on the experience from previous Landsat data studies (Li et al., 2022). But in fact, the NDVIv value is affected by the vegetation type changes as well as temporal changes (Xia et al., 2017). Besides, NDVIs will take different values due to differences in soil color, composition, moisture content, roughness and other factors, with a range that spans from -0.1 to 0.2 (Rundquist, 2002). Therefore, the NDVI values of pure pixels obtained empirically cannot be applied directly, but should be determined more precisely by further studies.
In this study, the DP models with [90%], [95%], [99%], [0.5%,99.5%], and [0.05%,99.95%] as the value intervals of [NDVIs, NDVIv] were established. The differences in accuracy between the extracted vegetation cover and the true value under each model were calculated to determine the optimal model. According to this calculation, the optimal models for the other VIs could also be attained. The DN values for each VI at each interval range were determined using ENVI Classic 5.3 software (Table 3).
Table 3 DN value statistics based on the value intervals of [NDVIs, NDVIv]
DN value [10%, 90%] [5%, 95%] [1%, 99%] [0.5%, 99.5%] [0.05%, 99.95%]
NDVI [0.051322, 0.133003] [0.040184, 0.155279] [0.02162, 0.203545] [0.010482, 0.225821] [‒0.030358, 0.34463]
SAVI [‒0.004837, 0.182795] [‒0.005689, 0.216203] [‒0.006371, 0.283018] [‒0.006456, 0.316426] [‒0.034355, 0.477896]
MSAVI [‒0.005421, 0.215431] [‒0.006424, 0.2482] [‒0.007227, 0.320291] [‒0.007227, 0.346506] [‒0.046718, 0.484134]
ARVI [‒0.092250, ‒0.009038] [‒0.098194, 0.008793] [‒0.116025, 0.056343] [‒0.121969, 0.092006] [‒0.1398, 0.228712]
MVI [0.000958, 0.350775] [0.000479, 0.378963] [0.0000958, 0.435337] [0.00004791, 0.460393] [0.000004791, 0.566878]
EVI [‒0.009418, 0.309804] [‒0.01062, 0.364706] [‒0.01158, 0.513725] [‒0.0117, 0.631373] [‒1, 0.992157]

2.3.2 Multi-VIs model

(1) VI Multicollinearity diagnosis
In regression studies, the phenomenon of linear relationships or high correlations among independent variables is multicollinearity. The existence of multicollinearity among data often leads to problems such as the uniqueness of regression solutions, increases in the standard errors of parameter estimates and reductions in model stability (Sun, 2020). Therefore, multicollinearity among variables should be diagnosed as the first step in regression studies.
Multicollinearity among variables is generally diagnosed by the variance inflation factor (VIF), which is an indicator that measures how much the multicollinearity among explanatory variables enhances the variance of the estimated coefficients. The larger the value of VIF, the more serious the degree of covariance. Empirical studies have determined that when VIF>10, the covariance is severe. The VIF formula is as follows (Chen and Wang, 1984):
$VI{{F}_{i}}=\frac{1}{1-R_{i}^{2}}$
where $VI{{F}_{i}}$ denotes the variance inflation factor of the i-th regression coefficient, and $R_{i}^{2}$ denotes the complex correlation coefficient of the regression of independent variable i on all of the other independent variables except for i.
The VIF function in RStudio software was used to calculate the variance inflation factors between variables, and the results are shown in Table 4.
Table 4 VIF values between VIs
VIF values Remaining variables
SAVI MSAVI ARVI EVI MVI
NDVI 14859.03 24919.92 251.70 408.94 1566.37
SAVI 276.82 250.61 408.75 159.06
MSAVI 248.70 383.12 39.92
ARVI 16.30 16.30
EVI -
According to Table 4, strong or even very strong multicollinearity exists among the indexes. Therefore, the commonly used regression models based on least squares estimation of regression coefficients are not applicable, and models that can solve or avoid multicollinearity among the data should be selected to obtain more accurate inversion effects.
Based on these considerations, three models were chosen in this study: R regression, PCA regression and RF regression models.
(2) Ridge regression model (R model)
In contrast to the traditional estimation of regression coefficients using least squares, Ridge regression (R model) is a biased estimation method based on least squares that is specifically designed to deal with data covariance (Zhao et al., 2019). It has high stability and can solve the problem of the model over-fitting phenomenon pretty well (Yuan et al., 2019).
The core idea is to assume that the model is$y=X\beta +\varepsilon $, and then reduce the singularity of ${X}'X$ by giving the addition matrix kI (k>0) of ${X}'X$ in the original least squares estimate (Fang, 1988). The estimated value is shown in Equation (10).
$\hat{\beta }$(k)=${{\left( {X}'X+kI \right)}^{-1}}{X}'$y
where k is the R coefficient, $\hat{\beta }$ (k) is the parameter estimate and ${X}'$ is the transpose matrix of X.
The ridge coefficients can be ascertained by plotting the ridge trace curves, i.e., the relationship curves of the independent variables with the ridge estimation parameters, of the ridge trace method. The ridge coefficient can be selected when all the curves become stable. For this study, the ridge trace curve is shown in Fig. 4, and when k is equal to 2.3, all the parameters tended to stabilize. So the k value of 2.3 is used.
Fig. 4 Ridge trace curves
(3) Principal component analysis model (PCA model)
Principal component analysis (PCA) mainly applies the idea of dimensionality reduction to transform multiple original variables into fewer mutually independent composite variables by orthogonal rotation with less information loss in the original data (Zhao, 2014). These composite variables are the principal components. Principal component regression is the process of determining the relationship between the original variables and the target variables by considering the extracted principal components as intermediate variables. This model eliminates multicollinearity in the regression model because the principal components are uncorrelated with each other (Yu, 2006).
The PCA regression model was implemented through the princomp function in RStudio software. According to the gravel plot (Fig. 5) with the cumulative variance contribution (Table 5), the number of principal components of the PCA model is 1.
Fig. 5 Gravel map of the PCA model
Table 5 Cumulative variance contributions of the principal components
Principal components Feature root Contribution rate (%) Cumulative
contribution rate (%)
1 2.4303820 98.4459 98.4459
2 0.2863041 1.3662 99.8121
3 0.0987407 0.1625 99.9746
(4) Random forest regression model (RF model)
The random forest regression model (RF model) is a process of randomly sampling data sets and variables, constructing regression decision trees with different rules for different attributes based on each sampled set, and then using the mean value of these regression decision tree results as the final result (Xing et al., 2021). These decision trees are independent of each other, so RF regression is also capable of overcoming the multicollinearity among the data (Breiman, 2001).
The RF regression model was constructed using the randomForest package in the RStudio software. Two important parameters of this model, the number of variables available at each mode (mtry) and the decision tree (Ntree), were obtained by plotting the parameters versus the error to form the model (Fig. 6). The error basically tended to be stable when Ntree was equal to 300 and the error value was minimized when mtry was equal to 4. Therefore, the Ntree value of this model was set as 300 while the mtry value was 4.

2.3.3 Evaluation of model accuracy

The evaluation of the FVC inversion model accuracy was performed by calculating the accuracy evaluation statistics, so the Residual Sum of Squares (SSE) and Coefficient of Determination (R2) of the training set, and the Root Mean Square Error (RMSE) of the test set were also calculated. When R2 is closer to 1 and the SSE is smaller, then the model fitting accuracy is higher. At the same time, the smaller the RMSE, the higher the model prediction accuracy. The calculation formula is:
$SSE=\underset{i=1}{\overset{n}{\mathop \sum }}\,{{\left( {{y}_{i}}-\hat{y} \right)}^{2}}$
${{R}^{2}}=1-\frac{\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\hat{y} \right)}^{2}}}}{\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\hat{y} \right)}^{2}}}}$
$RMSE=\sqrt{\frac{\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\hat{y} \right)}^{2}}}}{n}}$
Fig. 6 Significant variables of the decision tree model
Where ${{y}_{i}}$ is the measured value; $\hat{y}$ is the model predicted value; $\bar{y}$ is the mean of the measured value; and n is the number of samples.
In this study, the 77 sets of field survey data were divided into a training set with 52 data points and a test set with 25 data points by using the createDataPartition function of RStudio software with a ratio of about 2:1. The training set data was used for modeling while the test set was used for model accuracy testing.
Overfitting is a very common phenomenon during model training, i.e., the model performs well in the training set but not in the test set. To verify whether there is overfitting of the model, or whether the model has a good generalization ability, the k-fold cross-validation method is often used. The so-called k-fold cross-validation is the process of dividing the data into k subsets that are mutually exclusive and approximately equal in data size, then using k-1 of these subsets for training and the remaining one for testing, and then calculating the accuracy of each training model separately.

3 Results and analysis

3.1 Single VI model accuracy evaluation

The DP model was built for each VI on different intervals and the accuracy statistics were calculated (Fig. 7). The results in Fig. 7 intuitively reflect that the accuracy increases with the expansion of the interval for both MVI and MSAVI, so [0.05%,99.95%] is the highest accuracy interval. The accuracy of the NDVI, SAVI, EVI and ARVI models all follow a trend of decreasing and then increasing with the expansion of the interval. Thus, the highest accuracy intervals of the [$\text{NDV}{{\text{I}}_{\text{s}}}$, $\text{NDV}{{\text{I}}_{\text{v}}}$] are [0.5%,99.5%] for NDVI, SAVI and EVI, and [99%] for ARVI. Among the six VIs, NDVI has the smallest volatility amplitude, so it can be considered as the model for which the accuracy is least affected by the interval range.
Fig. 7 Evaluation of the accuracy of the DP model
Overall, the precision of MVI was lower than all the other VIs under the optimal interval, so the DP model developed by MVI is not applicable to FVC extraction in the study area. The statistics of the remaining VIs under their optimal intervals were collated (Table 6) for further comparative analysis.
Table 6 The statistics of five VIs (excluding MVI) under their optimal intervals
VIs with
optimal interval		 NDVI
$\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]$		 SAVI
$\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]$		 MSAVI
$\left[ 0.05\text{ }\!\!%\!\!\text{ },\ 99.95\text{ }\!\!%\!\!\text{ } \right]$		 ARVI
$\left[ 1\text{ }\!\!%\!\!\text{ },\ 99\text{ }\!\!%\!\!\text{ } \right]$		 EVI
$\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]$
SSE 0.4053296 0.2824413 0.3928578 0.3657897 0.2897076
R2 0.672436 0.7717473 0.682515 0.7043899 0.765875
RMSE 0.2781874 0.0656734 0.067913 0.0737328 0.0089758
According to Table 6, the sequence of error values of the five VIs (i.e., the sum of SSE and RMSE after unifying the dimensionals) is NDVI>MSAVI>ARVI>SAVI>EVI. The sequence for the comparison of R²values is SAVI>EVI> ARVI>MSAVI>NDVI. These results indicate that among the six VIs DP models, the models established by SAVI $\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]$ and EVI $\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]$ are the most suitable for FVC extraction in the study area.

3.2 Multi-VI model accuracy evaluation

3.2.1 Model accuracy statistics

The prediction results of the R, PCA and RF models were plotted as scatter plots (Fig. 8), and the statistics for the accuracy of each model were calculated (Fig. 9).
Fig. 8 Model prediction scatter plots for the PCA, R and RF models
Fig. 9 Model accuracy evaluation of the PCA, R and RF models
In the model prediction scatter plots (Fig. 8), the straight lines indicate the cases where the estimated values are equal to the measured values, and the closer the scatter points are to the straight line, the higher the estimation accuracy. Among the three models, about two-thirds of the PCA model training set points (Fig. 8a) fall above the straight line, indicating a possible overestimation. The scatter of the points for the R and RF models are basically evenly distributed on both sides of the straight line. There might be an underestimation in the areas with high coverage and overestimation in areas with low coverage, although there is not enough evidence to tell whether this underestimation is accidental due to the low data distribution in the areas with high FVC.
Combining the numerical result statistics of the three types of model in Fig. 9 clearly shows the three relationships of: $SS{{E}_{RF}}<SS{{E}_{R}}<SS{{E}_{PCA}}$, $R{{{}^\text{2}}_{\text{PCA}}}<R{{{}^\text{2}}_{\text{R}}}<R{{{}^\text{2}}_{\text{RF}}}$ and $\text{RMS}{{\text{E}}_{\text{R}}}<$$\text{RMS}{{\text{E}}_{\text{RF}}}<\text{RMS}{{\text{E}}_{\text{PCA}}}$. Therefore, the RF model fitted the training set better than the other two models, the R model had the highest prediction accuracy, and the PCA model performed poorly in both model fitting and prediction. Since the R and RF models were dominant in different aspects separately, directly picking the better model is not possible. Meanwhile, the statistical results of the PCA model were similar to those of the R model, so there is no way to judge whether this superior or inferior result was accidental. Therefore, other methods are needed for further comparisons.

3.2.2 Model stability evaluation: K-fold cross verification

In order to further compare the applicability of the three models in the study area from the perspective of stability, the three models were tested using the K-fold cross-validation method, where the K value was set to 5 so the data were randomly split into five groups. The data division was implemented through the createFolds function of RStudio software. The statistics were calculated for the models built when each of the four data sets was used as the test set while the remaining data were the training set (Fig. 10), and the means of these statistics were also calculated (Table 7).
Fig. 10 Five-fold cross-validation results
Table 7 Means of the cross-validation results
Model $\overline{\text{SSE}}$ ${{\bar{R}}^{2}}$ $\overline{\text{RMSE}}$
PCA 0.364548 0.712160 0.050014
R 0.347112 0.724601 0.057356
RF 0.084036 0.933460 0.056530
Figure 10 and Table 7 show the results of the five-fold cross-validation of the three models. Considering only the fitting effect of the models (SSE and R2), the RF model was superior to the R and PCA models (Fig. 10a) and the PCA had the lowest accuracy. Considering only the prediction effect (RMSE), the PCA model had the smallest fluctuation amplitude in Fig. 10c and its mean value of RMSE was the lowest (Table 7), which indicated that it was more stable. However, between the R and RF models, ${{\overline{\text{RMSE}}}_{\text{RF}}}<$ ${{\overline{\text{RMSE}}}_{\text{R}}}$, which indicated that the RF model was more stable than R model. Then combining these findings with the original model accuracy statistics results, although the R model possessed the best prediction ability, it also had strong instability. The PCA model was extremely stable, whereas, it had low accuracy. In comparison, the RF model had a certain level of stability while expressing good precision, consequently, the RF model is the most suitable overall for extracting the FVC among these models.

3.3 Comprehensive evaluation of single index and multi-index models

The statistics of the models screened in Sections 3.1 and 3.2 are listed in Table 8 for comparison. The results clearly show that the RF model demonstrated a strong fitting ability, and the DP model composed of EVI had the highest prediction precision. Owing to the distinctiveness of the modeling mechanisms of the DP model and the RF model, that is, the RF model was established based on the measured data and the DP model is based on remote sensing imagery, these two models were not compared further. Ultimately, the DP model constituted by EVI and the RF model of the six VIs were selected from among the various models to extract the FVC of the study area.
Table 8 Statistics of the single VI & muti-VI models
Model type VI SSE R2 RMSE
DP model $\text{SAV}{{\text{I}}_{\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]}}$ 0.282441 0.771747 0.065673
DP model $\text{EV}{{\text{I}}_{\left[ 0.5\text{ }\!\!%\!\!\text{ },\ 99.5\text{ }\!\!%\!\!\text{ } \right]}}$ 0.289708 0.765875 0.008976
RF model Six VIs 0.049816 0.959742 0.018689

3.4 FVC extraction for the study area

The DP model of EVI and RF model were both utilized to extract the FVC values of the whole study area.
(1) DP model (EVI)
The result of FVC extraction by applying the DP model constructed using the EVI is shown in Fig. 11.
Fig. 11 The FVC result of the study area (DP model based on EVI)
(2) RF model
Applying all measured data to the modeling of the RF model, two parameters of the model (Ntree and mtry) were ascertained on the basis of step 2.3.4 above. The results of FVC extraction were then plotted as a raster map (Fig. 12). Since the RF model modeling with a non-physical structure is implemented in RStudio software, the drawing of the extracted plot was completed by randomly selecting 100000 sample points combined with kriging interpolation.
Fig. 12 The FVC result of the study area (RF model)
(3) The FVC of the study area
The distributions of FVC in the two maps in Figs. 11 and 12 are basically similar, which may further support the applicability of both models.
The FVC of the whole study area tends to be low, while high vegetation coverage blocks are mainly concentrated in the south. The overall distribution is similar to the texture of the mountains, which may indicate that the distribution of vegetation cover in the study area is closely related to topographic factors. The correlation between the two extraction results was calculated and a strong correlation was obtained (r²=0.82, P<0.1, two-tailed Pearson test), indicating that using the two models to estimate vegetation coverage in the study area is feasible. Intuitively, however, the figures show a large difference in the two extraction diagrams. Two possible reasons for these differences are that they may represent an error caused by the selection of the drawing method or they may come from the result itself. In order to improve the rationality of the results, the mean of these two models can be used as the actual application results.

4 Estimation errors and uncertainty analysis

The sources of research error may include matching error, actual measurement error and image error.
First of all, the accuracy verification process will generate errors when locating the actual sample points from remote sensing images. On the one hand, although the GPS positioning with higher accuracy (greater than 1 m) was chosen in the actual measurement process, some of these errors will still exist. One the other hand, the size of the measured sample is 1 m×1 m, which is exactly equal to 1×1 pixel of the GF-2 image, so errors may be generated if the sample vertices do not match exactly with the pixels. However, the vegetation in the study area is evenly distributed, which means that the small differences between the sample points and the surrounding area can effectively reduce these errors.
Secondly, the visual estimation method may cause errors due to human subjective influences. These errors mainly originate from the visual estimation itself and the inconsistent standards of individual estimations. The former is more difficult to avoid, but it can be reduced by calculating the sample FVC in blocks. The latter can be circumvented by having all sample points counted by the same person. In this study, the error was minimized as much as possible by following these procedures.
Finally, during the image extraction process, the image may be shifted and distorted due to wind, clouds, solar altitude angle, sensors and other factors, thus creating errors. However, these problems are largely corrected by the radiation calibration, atmospheric correction and orthorectification correction in the preprocessing step.

5 Conclusions

In order to achieve high accuracy and fast extraction of FVC in the low disturbance area of the Northwest Arid Desert, and thus provide reference for vegetation restoration in mining areas, this study combined GF-2 remote sensing images with the actual measurement data to screen the models for estimating FVC. First, we pre-processed the GF-2 images, including radiometric calibration, atmospheric correction, orthorectification and image fusion. Then the processed images were combined with measured sample point data to extract the VI data for training the dataset with PCA, R and RF regression models separately. Through the above procedures, the screening of remote sensing estimation models of vegetation cover in the low disturbance areas of arid deserts was realized. Eventually, the selected model was applied to invert the FVC of the entire study area. This analysis led to four main conclusions.
(1) The accuracy of the DP models constructed by individual VIs will be affected by the values of pure vegetation and bare soil pixels ($V{{I}_{v}}$ and $V{{I}_{s}}$) to a large extent. However, compared to the other five VIs, NDVI is less affected by them. Besides, the EVI model shows both good fitting accuracy and excellent prediction accuracy, so it was selected as the most preferable DP model for FVC extraction in this research.
(2) Among the multi-VI models, the RF model has superior accuracy in fitting and prediction, and it shows a certain stability. So, it was also picked for extraction.
(3) Due to their different modeling principles, these two models were not be further compared. The extracted results have a strong correlation, indicating that applying these two models for extracting FVC in the study area is feasible. The results also show that the RF model has good generalizability and scalability. In practical applications, the average of the two models can be used as the final result.
(4) The extraction results also show that the vegetation distribution in the study area is similar to the outline of the mountain range, so it could be inferred that the topographic distribution in the area is the dominant factor driving the vegetation distribution. Therefore, subsequent studies should conduct more in-depth research in this area in conjunction with elevation maps. In addition, vegetation growth is a dynamic process, so it will be necessary to carry out follow-up research on the spatial and temporal distribution patterns of vegetation in the study area.

Acknowledgements

Thanks to the data support provided by Beijing Shenghaiyanuo Information Technology Co.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Bannari A, Morin D, Bonn F, et al. 1995. A review of vegetation indices. Remote Sensing Reviews, 13: 95-120.

DOI

[2]
Breiman L. 2001. Random forests. Machine Learning, 45: 5-32.

DOI

[3]
Chen H, Lei S G. 2018. Artificial vegetation extraction for mining area based on object-oriented techniques and the analysis of reconstruction. Journal of Arid Land Resources and Environment, 32(7): 98-103. (in Chinese)

[4]
Chen Q, Li X S, Xiu X M, et al. 2019. Large scale shrub coverage mapping of sandy land at 30m resolution based on Google Earth Engine and machine learning. Acta Ecologica Sinica, 39(11): 4056-4069. (in Chinese)

[5]
Chen X R, Wang S G. 1984. Modern practical regression analysis. Nanning, China: Guangxi People’s Publishers. (in Chinese)

[6]
Cui W X, Li J R, Si Q C, et al. 2021. Research on extraction method of desert shrub coverage based on UAV visible light data. Research of Soil and Water Conservation, 28(6): 175-182, 189. (in Chinese)

[7]
Deng Z D, Lu Z, Wang G Y, et al. 2021. Extraction of fractional vegetation cover in arid desert area based on Chinese GF-6 satellite. Open Geosciences, 13(1): 416-430.

DOI

[8]
Ding Y L. 2015. Research on the estimation of fractional vegetation cover and the validation of fractional vegetation cover product. Diss., Beijing, China: University of Chinese Academy of Sciences. (in Chinese)

[9]
Fang K T. 1988. Practical regression analysis. Beijing, China: Science Publishers. (in Chinese)

[10]
Feng H Y, Feng Z K, Feng H X. 2017. A new method for estimating the fractional vegetation cover based on UVA hyperspectrum. Spectroscopy and Spectral Analysis, 37(11): 3573-3578. (in Chinese).

[11]
Gao Y G, Xu H Q. 2017. Estimation of multi-scale urban vegetation coverage based on multi-source remote sensing images. Journal of Infrared and Millimeter Waves, 36(2): 225-234. (in Chinese)

[12]
Gao Y P, Kang M D, He M Z, et al. 2018. Extraction of desert vegetation coverage based on visible light band information of unmanned aerial vehicle: A case study of Shapotou region. Journal of Lanzhou University (Natural Sciences), 54(6): 770-775. (in Chinese)

[13]
Gu X H, Zhang Y, Sang H Y, et al. 2020. Research on crop classification method based on Sentinel-2 time series combined vegetation index. Remote Sensing Technology and Application, 35(3): 702-711. (in Chinese)

[14]
Gutman G, Ignatov A. 1998. The derivation of the green vegetation fraction from NOAA/AVHRR data for use in numerical weather prediction models. International Journal of Remote Sensing, 19(8): 1533-1543.

DOI

[15]
Huete A R. 1988. A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25(3): 295-309.

DOI

[16]
Ji C C. 2018. Reasarch on muti-scale spectral mixture analysis method for sparse photosynthetic/non-photosynthetic vegetation in arid area. Diss., Wuhan, China: Wuhan University. (in Chinese)

[17]
Jia K, Yao Y J, Wei X Q, et al. 2013. A review on fractional vegetation cover estimation using remote sensing. Advances in Earth Science, 28(7): 774-782. (in Chinese)

DOI

[18]
Kaufman Y J, Tanre D. 1992. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Transactions on Geoscience and Remote Sensing, 30(2): 261-270.

DOI

[19]
Li M M. 2003. The method of vegetation fraction estimation by remote sensing. Diss., Beijing, China: Graduate School of Chinese Academy of Sciences (Institute of Remote Sensing Application). (in Chinese)

[20]
Li X, Chen J L, Deng S Q. 2022. Long time series detection of vegetation coverage change based on Google Earth Engine. Beijing Surveying and Mapping, 36(4): 457-462. (in Chinese)

[21]
Li Y Y, Song Y Y, Jiang X L. 2015. Study on the differences among 3 kinds of vegetation indexes (Ⅵ) in extracting coverage of desert forest with remote sensing image. Journal of Shihezi University (Natural Science), 33(1): 72-77. (in Chinese)

[22]
Liao K T, Song Y J, Xie S H, et al. 2022. Spatial and temporal variation characteristics of vegetation coverage in Jiangxi Province based on Google Earth Engine. Soil and Water Conservation in China, (10): 64-67, 9. (in Chinese)

[23]
Liu L, Zhen X M, Jiang T, et al. 2021a. Extraction method of fractional vegetation cover from unmanned aerial vehicle remote sensing: A review. Journal of Northeast Normal University (Natural Science Edition), 53(4): 151-160. (in Chinese)

[24]
Liu Y, Xu L, Feng F, et al. 2021b. Species diversity and community stability at early stage of vegetation reclamation in gangue hill slope of Wuhai Mining Area. Bulletin of Soil and Water Conservation, 41(1): 190-196. (in Chinese)

[25]
Luo Y, Xu J H, Yue W Z. 2005. Research on vegetation indices based on the remote sensing images. Ecological Science, 24(1): 75-79. (in Chinese)

[26]
Lv L L. 2018. Sparse vegetation extraction and change analysis based on spectral differentiation: An example in the Shule River Basin Plain. Diss., Lanzhou, China: Lanzhou University. (in Chinese)

[27]
McDaniel K, Haas R H. 1982. Assessing mesquite-grass vegetation condition from Landsat. Photogrammetric Engineering and Remote Sensing, 48: 441-450.

[28]
Meng D C, Zhao J, Lan Y B, et al. 2020. Vegetation coverage extraction model of winter wheat based on pixel dichotomy. Journal of South China Agricultural University, 41(3): 126-132. (in Chinese)

[29]
Pi X Y, Zeng Y N, He C Q. 2021. High-resolution urban vegetation coverage estimation based on multi-source remote sensing data fusion. National Remote Sensing Bulletin, 25(6): 1216-1226. (in Chinese)

DOI

[30]
Qi J, Chehbouni A, Huete A R, et al. 1994. A modified soil adjusted vegetation index. Remote Sensing of Environment, 48(2): 119-126.

DOI

[31]
Ren C S, Ye H C, Cui B, et al. 2017. Acreage estimation of mango orchards using object-oriented classification and remote sensing. Resources Science, 39(8): 1584-1591. (in Chinese)

DOI

[32]
Rouse J W, Hass R H, Schell J A, et al. 1973. Monitoring vegetation systems in the Great Plains with ERTS. Third Earth Resources Technology Satellite-1 Symposium. Washington DC, USA: The National Aeronautics and Space Administration.

[33]
Rundquist B C. 2002. The influence of canopy green vegetation fraction on spectral measurements over native tallgrass prairie. Remote Sensing of Environment, 81(1): 129-135.

DOI

[34]
Shen Q, Zhu C M, Zhang X. 2019. Fractional vegetation cover estimation in aird and rare vegetation area aided by GF-2 remote sensing data. Bulletin of Surveying and Mapping, (7): 33-38. (in Chinese)

DOI

[35]
Su Y, Qi Y, Wang J H, et al. 2017. Vegetation coverage classification and vegetation structure parameters extraction of Populus euphratica forest in Ejina Oasis by LiDAR data. Journal of Desert Research, 37(4): 689-697. (in Chinese)

[36]
Su Y Q, Lai R W, Yan Q, et al. 2018. Extraction of vegetation coverage and analysis of landscape pattern. Journal of Forest and Environment, 38(2): 164-170. (in Chinese)

[37]
Sun J C. 2020. Ridge estimation method for multicollinearity in linear regression models. Diss., Jinzhou, China: Bohai University. (in Chinese)

[38]
Tan C, Wang F, Wang N N, et al. 2022. Extraction of vegetation coverage in red line area of ecological protection by using HJ-2A/B satellites images. Spacecraft Engineering, 31(3): 153-158. (in Chinese)

[39]
Tian J R, Dai T T, Xu Y N, et al. 2018. Extraction of vegetation parameters in different stages of ecological restoration on abandoned mine area based on T-LiDAR. Journal of Ecology and Rural Environment, 34(8): 686-691. (in Chinese)

[40]
Wang H. 2021a. Estimation method and spatio-temporal continuous mapping of fraction vegetation cover with Landsat data. Diss., Wuhan, China: Wuhan University. (in Chinese)

[41]
Wang J Q. 2021b. Extraction and analysis of vegetation coverage in the source region of the Yellow River based on the Multiple Endmember Spectral Mixture Analysis. Diss., Beijing, China: China University of Geosciences. (in Chinese)

[42]
Wang X X, Lu X P, Li G Q, et al. 2022a. Combining the red edge-near infrared vegetation indexes of DEM to extract urban vegetation information. Spectroscopy and Spectral Analysis, 42(7): 2284-2289. (in Chinese)

[43]
Wang Y F, Tan L, Guo H Li, et al. 2022b. Study on spatial and temporal variation of forest fractional vegetation cover in Liyang based on multi-source remote sensing data. Journal of Nanjing Forestry University (Natural Sciences), (40): 1-12. (in Chinese)

[44]
Wang Z, Zhao L J, Niu K, et al. 2021. A review of vegetation cover extraction methods based on remote sensing images. Agriculture and Technology, 41(14): 25-29. (in Chinese)

[45]
Wu Q Y, Gao F, Li Z X, et al. 2021. The contrast of Sentinel-2A and Landsat 8 data in remote sensing estimation of vegetation coverage. Bulletin of Surveying and Mapping, (S1): 104-108, 113. (in Chinese)

[46]
Xia Y, Fan J R, Li L L, et al. 2017. Comparisons on sparse alpine grassland based on vegetation coverage inversion models of remote sensing. Journal of Sichuan Agricultural University, 35(1): 37-44, 59. (in Chinese)

[47]
Xing X Y, Yang X C, Xu B, et al. 2021. Remote sensing estimation of grassland aboveground biomass based on random forest. Journal of Geo-Information Science, 23(7): 1312-1324. (in Chinese)

[48]
Ye J Y. 2017. Remote sensing information extraction and inversion of vegetation in dryland area. Diss., Beijing, China: Chinese Academy of Forestry. (in Chinese)

[49]
Yu C H. 2006. SPSS statistical analysis. Wuhan, China: Electronic Industry Publishers. (in Chinese)

[50]
Yuan C, Xiang L Y, Bai R, et al. 2019. MiR-195 restrains lung adenocarcinoma by regulating CD4+ T cell activation via the CCDC88C/Wnt signaling pathway: A study based on the Cancer Genome Atlas (TCGA), Gene Expression Omnibus (GEO) and bioinformatic analysis. Annals of Translational Medicine, 7(12): 263. DOI: 10.1016/j.jtho.2019.08.987.

DOI PMID

[51]
Yue X X, Liu G H, Ren L, et al. 2011. Floristic analysis of spermatophytes in the steppe desert area of Wuhai. Acta Agrestia Sinica, 19(1): 7-12. (in Chinese)

[52]
Zhang H B. 2022. Extraction and temporal and spatial distribution characteristics of aquatic vegetation in typical lakes in Northern China. Diss., Changchun, China: Jilin University. (in Chinese)

[53]
Zhang X L, Feng X. 2022. Research and analysis of vegetation cover change in Yunnan Province based on NDVI. Modern Agriculture, (10): 63-66. (in Chinese)

[54]
Zhao J J, Lu M X, Gu H H, et al. 2018. Research on ecological restoration effect of mining area based on vegetation cover change. Mining Research and Development, 38(10): 115-118. (in Chinese)

[55]
Zhao J S, Li W, Wang W S, et al. 2021. Analysis on saturation of Suaeda salsa vegetation index based on DART model. Journal of Central China Normal University (Natural Sciences), 55(1): 121-127. (in Chinese)

[56]
Zhao L L, Zhang R, Liu Y X, et al. 2020. The differences between extracting vegetation information from GF1-WFV and Landsat8-OLI. Acta Ecologica Sinica, 40(10): 3495-3506. (in Chinese)

[57]
Zhao Y, Zhang L, Zhang Y, et al. 2019. Identification of hedgehog signaling as a potential oncogenic driver in an aggressive subclass of human hepatocellular carcinoma: A reanalysis of the TCGA cohort. Science China (Life Sciences), 62(11): 1481-1491.

[58]
Zhao T A. 2014. The study of vegetation coverage modeling and dynamic changes in Miyun County. Diss., Beijing, China: Beijing Forestry University. (in Chinese)

[59]
Zheng S Y, Hai Y, He M Q, et al. 2022. Comparative study on vege tation extraction effect based on ZY-102D data. Spacecraft Recovery & Remote Sensing, 43(2): 92-103. (in Chinese)

[60]
Zhou T, Hu Z Q, Han J Z, et al. 2021. Green vegetation extraction based on visible light image of UAV. China Environmental Science, 41(5): 2380-2390. (in Chinese)

Options
Outlines

/

〈 〉