Estimation of Photosynthetic and Non-Photosynthetic Vegetation Coverage in the Lower Reaches of Tarim River Based on Sentinel-2A Data

Abstract

Estimating the fractional coverage of the photosynthetic vegetation (f_PV) and non-photosynthetic vegetation (f_NPV) is essential for assessing the growth conditions of vegetation growth in arid areas and for monitoring environmental changes and desertification. The aim of this study was to estimate the f_PV, f_NPV and the fractional coverage of the bare soil (f_BS) in the lower reaches of Tarim River quantitatively. The study acquired field data during September 2020 for obtaining the f_PV, f_NPV and f_BS. Firstly, six photosynthetic vegetation indices (PVIs) and six non-photosynthetic vegetation indices (NPVIs) were calculated from Sentinel-2A image data. The PVIs include normalized difference vegetation index (NDVI), ratio vegetation index (RVI), soil adjusted vegetation index (SAVI), modified soil adjusted vegetation index (MSAVI), reduced simple ratio index (RSR) and global environment monitoring index (GEMI). Meanwhile, normalized difference index (NDI), normalized difference tillage index (NDTI), normalized difference senescent vegetation index (NDSVI), soil tillage index (STI), shortwave infrared ratio (SWIR32) and dead fuel index (DFI) constitutes the NPVIs. We then established linear regression model of different PVIs and f_PV, and NPVIs and f_NPV, respectively. Finally, we applied the GEMI-DFI model to analyze the spatial and seasonal variation of f_PV and f_NPV in the study area in 2020. The results showed that the GEMI and f_PV revealed the best correlation coefficient (R²) of 0.59, while DFI and f_NPV had the best correlation of R² = 0.45. The accuracy of f_PV, f_NPV and f_BS based on the determined PVIs and NPVIs as calculated by GEMI-DFI model are 0.69, 0.58 and 0.43, respectively. The f_PV and f_NPV are consistent with the vegetation phonological development characteristics in the study area. The study concluded that the application of the GEMI-DFI model in the f_PV and f_NPV estimation was sufficiently significant for monitoring the spatial and seasonal variation of vegetation and its ecological functions in arid areas.

1. Introduction

Vegetation cover is an important part of the ecosystem, which plays an important role in the soil and water conservation, restraining the desertification process, biodiversity protection and other ecological service and functions etc. [1,2]. From a functional point of view, vegetation can be divided into two parts: the photosynthetic vegetation (PV) and the non-photosynthetic vegetation (NPV) [3]. The PV mainly refers green leaves, while the NPV mainly includes the vegetation litter (dead leaves, dead branches and stems) and crop residues [4,5]. NPV is important as it not only affects the carbon storage, CO₂ exchange capacity and temperature between the surface and soil [6,7] but also slows down soil erosion, increases the soil organic matter and improves the soil quality [8,9,10,11,12]. Therefore, an accurate estimation of the fractional coverage of the photosynthetic (f_PV) and non-photosynthetic vegetation (f_NPV) is of great significance in evaluating the ecological conditions.

In recent decades, remote sensing technology has made great progress in many aspects of geography [13,14]. In the f_PV estimation, a number of indices (e.g. normalized difference vegetation index (NDVI) [15], soil adjusted vegetation index (SAVI) [16] and global environment monitoring index (GEMI) [17]) have been developed for PV utilizing spectral reflectance, which is widely used in various satellite platforms. Researchers have also proposed some vegetation indices so as to estimate the f_NPV [18,19,20,21,22,23]. According to the spectral resolution of remote sensing data, non-photosynthetic vegetation indices (NPVIs) can be divided into hyperspectral NPVIs and multispectral NPVIs. Daughty [21] proposed hyperspectral NPVI (cellulose absorption index (CAI)) based on hyperspectral reflection characteristics of the NPV. Taking into account the limitations of hyperspectral data, researchers have developed many multispectral NPVIs using Landsat TM bands, such as the normalized difference index (NDI), the normalized difference tillage index (NDTI), the normalized difference senescent vegetation index (NDSVI) and the soil tillage index (STI). Guerschman [3] developed the shortwave infrared ratio index (SWIR32) based on MODIS data, which accurately estimated the f_NPV of grasslands in Australia. Cao [24] utilized multispectral, MODIS, data to estimate NPV using the dead fuel index (DFI).

In order to estimate the proportion of PV and NPV in a pixel, based on spectral mixture analysis (SMA), researchers construct pixel unmixed model by using PVI and NPVI to estimate f_PV and f_NPV. And based on this, Guerschman [3] used the Hyperion data to propose pixel linear unmixed model of the NDVI-CAI (NDVI represents the PV, CAI represents the NPV). This updated methodology includes a better estimation of the spatiotemporal distribution of the PV and NPV in the sparse grasslands of Australia. Validating this methodology, Wang [25] applied the DFI index to construct the NDVI-DFI linear unmixed model based on the MOD09GHK surface reflectance data and effectively calculated the f_PV and f_NPV of the Xilingol grasslands. The foundational work mentioned above successfully showed suitable PV and NPV estimation using multispectral data but were limited in geographic and ecological scope. Focusing primarily on grasslands, other regions have not been explored yet. This work aims to estimate f_PV and f_NPV in riparian regions.

The Tarim River is a typical arid inland river basin. Due to the impact of human reclamation and irrigation in the upper reaches of the Tarim River, the water of the lower reaches has dropped sharply over the past 20 years, the Tetima Lake has dried up and the vegetation environment has deteriorated year by year. A comprehensive assessment of the vegetation changes in this area has been an ongoing effort. Some of the work, by Guli [26] and Li [27] employed a variety of models to discuss the suitability of the estimated vegetation coverage in the Tarim River Basin. Zhu [28] applied the method of transfer learning so as to extract and analyze the vegetation coverage changes in a long time series. These studies neglect to include NPV data in their analyses and rely largely on course resolution imagery, resulting in limited accuracy and uncertainty in the original estimations. Considering the improvements made to spectral instruments and the requirements for time and spatial resolution, Sentinel-2 satellite data are a good choice, specifically the multispectral imager (MSI) onboard could provide multispectral data with spatial resolution of 10 m and revisiting period of five days. Therefore, using the Sentinel-2 data with a high spatial and temporal resolution will be the focus to estimate the f_PV and f_NPV in the study area.

The objective of this study was to quantitatively estimate the seasonal variation in the f_PV, f_NPV and the fractional coverage of bare soil (f_BS)in the lower reaches of the Tarim River based on Sentinel-2A image data. The study is divided into the following sections: (1) select the best PVI and NPVI as estimation index of f_PV and f_NPV according to the correlation coefficient of the PVIs with f_PV and NPVIs with f_NPV by linear regression models analysis, (2) build pixel unmixing model based on Sentinel-2A image data and quantitatively evaluate the results of estimating f_PV and f_NPV by using field sampled data and (3) map different periods of f_PV, f_NPV and f_BS for the lower reaches of the Tarim River and analyze seasonal variation.

2. Materials and Methods

2.1. Study Area

The study area was located between the Dashkol Reservoir and Tetima Lake [39.5–40.59° N, 87.56–88.46° E] in the lower reaches of the Tarim River (Figure 1), which was surrounded by the Taklamakan Desert in the west and the Kuruktagh Desert in the northeast [29]. The study area’s annual precipitation ranged from 20 to 50 mm but the annual potential evaporation varied from 2500 to 3000 mm [30,31]. The total annual solar radiation varied in a range of 5692–6360 MJ m⁻² with an annual sunshine from 2780 to 2980 h [32]. The water supply of the vegetation in the area mainly depended on the water streamed from the upper reaches of the river. The vegetation was mostly distributed on the floodplain on the riverbank, which composed of trees, shrubs and herbs. Dominant trees included the Populus euphratica, Elaeagnus angustifolia, the shrubs consisted of the Tamarix ramosissima, Lycium ruthenicum, Halimodendron halodendron, and herbs Phragmites australis, Alhagi sparsifolia, Poacynum hendersonii, Karelinia caspia [2].

2.2. Datasets

2.2.1. Field Data

The fractional coverage samples of the PV, NPV and bare soil (BS) were collected from the Dashkol Reservoir, Gancaochang, Bozkol, Arghan, Kurgan and Tetima Lake. All samples were collected from 25 September to 29 September 2020 when the PV, NPV and BS existed simultaneously. Data collection and processing steps were as follows: we defined 10 m × 10 m squares, aligned to the north, and covered by a homogeneous vegetation distribution. The four corners and the center of the square were precisely located with GPS. Secondly, a smaller square of 1 m × 1 m was positioned randomly 3–5 times in the larger square. At the same time, a digital camera was used to take pictures vertically 1.6 m above the sample square. Each image was processed in ENVI 5.3. Photos were divided into NPV, PV and BS categories and training samples were established for each for supervised classification. Finally, we calculate the f_PV, f_NPV and f_BS of all 1 m × 1 m squares in each 10 m × 10 m square, and count the average value to represent the f_PV, f_NPV and f_BS in the 10 m × 10 m square. Figure 2 shows the field square acquisition and classification processing.

2.2.2. Remote Sensing Data

Sentinel-2A Level-1C data was used in this study. Images were from, seven periods on 5 April, 25 May, 24 July, 18 August, 12 September, 2 October and 6 November 2020. The data were radiometrically calibrated and geometrically corrected, downloaded from the ESA website (https://sci hub.copernicus.eu/home). The image preprocessing was carried out in SNAP 6.0 provided by ESA (Sentinel application platform) and the ENVI 5.3 software platform. The SNAP 6.0 Sen2Cor (Sentinel to Correction) plug-in was used to perform atmospheric correction processing on all L1C data, which resulted in 9 bands (Band 2, Band 3, Band 4, Band 5, Band 6, Band 7. Band 8a, Band 11 and Band 12) L2A surface reflectance data. And the 9-band data were combined into a single multiple image using composite bands in Quantum GIS (QGIS) and then resampled to 10 m pixels. Image data were mosaicked, and the study area was clipped. Water features were masked by using the modified normalized difference water index (MNDWI) combined with the threshold method (threshold value is −0.08).

2.3. Methods

The general workflow of this project was: (1) construct linear regression models of the PVIs and f_PV, NPVIs and f_NPV, respectively, to select the optimal PVIs and NPVIs, (2) analyze the feasibility of the selected PVIs and NPVIs so as to construct the response space, (3) apply the pixel linear unmixed model to determine the end member values of the PV, NPV and BS based on the Sentinel-2A image data, (4) quantitatively evaluate the estimation accuracy of the f_PV and f_NPV using the field sampled data, (5) map time series of the f_PV, f_NPV and f_BS for the lower reaches of the Tarim River and to analyze the seasonal variation. The flowchart is shown in Figure 3.

2.3.1. PVIs and NPVIs

PV was affected by chlorophyll and the cell structure and showed typical spectral characteristics of green vegetation, with obvious peaks and valleys [25]. In this study, we selected several PVIs for testing, including the NDVI [33], RVI [34], SAVI [35], MSAVI [16], RSR [36] and GEMI [17] (

Table 1. The Sentinel-2-based multispectral photosynthetic vegetation indices (PVIs) used for the photosynthetic vegetation (f_PV) estimation.

Vegetation Index	Equation	Citation
NDVI (normalized difference vegetation index)	$\mathrm{NDVI}=\frac{R_{\mathrm{NIR}}-R_{\mathrm{Red}}}{R_{\mathrm{NIR}}+R_{\mathrm{Red}}}$	Deering. 1978
RVI (ratio vegetation index)	$\mathrm{RVI}=\frac{R_{\mathrm{Red}}}{R_{\mathrm{NIR}}}$	Jordan. 1969
SAVI (soil adjusted vegetation index)	$\mathrm{SAVI}=\frac{R_{\mathrm{NIR}}-R_{\mathrm{Red}}}{R_{\mathrm{NIR}}+R_{\mathrm{Red}}+L}\left(1+L\right)$	Huete. 1988
MSAVI (modified soil adjusted vegetation index)	$\mathrm{MSAVI}=\frac{2\times R_{\mathrm{NIR}}+1-\sqrt{{\left(2\times R_{\mathrm{NIR}}+1\right)}^2-8\times \left(R_{\mathrm{NIR}}-R_{\mathrm{Red}}\right)}}{2}$	Qi et al., 1994
RSR (reduced simple ratio index)	$\mathrm{RSR}=\frac{R_{\mathrm{NIR}}}{R_{\mathrm{Red}}}\times \left(1-\frac{R_{\mathrm{SWIR}1}-R_{\mathrm{SWIR}1\min }}{R_{\mathrm{SWIR}1\max }-R_{\mathrm{SWIR}1\min }}\right)$	Brown et al., 2000
GEMI (global environment monitoringindex)	$\mathrm{GEMI}=\eta \times \left(1-0.25\eta \right)-\frac{R_{\mathrm{Red}}-0.125}{1-R_{\mathrm{Red}}}$; $\eta =2\times \frac{\left({R_{\mathrm{NIR}}}^2-{R_{\mathrm{Red}}}^2\right)+1.5\times R_{\mathrm{NIR}}+0.5\times R_{\mathrm{Red}}}{R_{\mathrm{NIR}}+R_{\mathrm{Red}}+0.5}$	Pinty et al., 1992

Notes: R_Red, R_NIR, R_SWIR1 and R_SWIR2 represent the reflectance of the Red band, Near Infrared band, Shortwave Infrared 1 band (1600 nm) and Shortwave Infrared 2 band (2100 nm), corresponding to the band 4, 8, 11 and 12 of sentinel-2A data, respectively. L is a soil adjustment factor, based on the experience L = 0.5.

).

NPV and BS had similar spectral reflectance characteristics in Visible and Near Infrared band and but could be differentiated using the Shortwave Infrared band (1600 nm and 2100 nm) [25]. In this study, we selected several PVIs for testing, including the NDI [23], NDTI [19], NDSAVI [20], STI [19], SWIR32 [3] and DFI [24] (

Table 2. The Sentinel-2-based multispectral non-photosynthetic vegetation indices (NPVIs) used for the non-photosynthetic vegetation (f_NPV) estimation.

Vegetation Index	Equation	Citation
NDI (normalized difference index)	$\mathrm{NDI}=\frac{R_{\mathrm{NIR}}-R_{\mathrm{SWIR}1}}{R_{\mathrm{NIR}}+R_{\mathrm{SWIR}1}}$	Mc Nairn et al., 1993
NDTI (normalized difference tillage index)	$\mathrm{NDTI}=\frac{R_{\mathrm{SWIR}1}-R_{\mathrm{SWIR}2}}{R_{\mathrm{SWIR}1}+R_{\mathrm{SWIR}2}}$	Deventer et al., 1997
NDSVI (normalized difference senescent vegetation index)	$\mathrm{NDSVI}=\frac{R_{\mathrm{SWIR}1}-R_{\mathrm{Red}}}{R_{\mathrm{SWIR}1}+R_{\mathrm{Red}}}$	Qi et al., 2002
STI (soil tillage index)	$\mathrm{STI}=\frac{R_{\mathrm{SWIR}1}}{R_{\mathrm{SWIR}2}}$	Deventer et al., 1997
SWIR32 (shortwave infrared ratio)	$\mathrm{SWIR}32=\frac{R_{\mathrm{SWIR}2}}{R_{\mathrm{SWIR}1}}$	Guerschman et al., 2009
DFI (dead fuel index)	$\mathrm{DFI}=100\times \left(1-\frac{R_{\mathrm{SWIR}2}}{R_{\mathrm{SWIR}1}}\right)\times \frac{R_{\mathrm{Red}}}{R_{\mathrm{NIR}}}$	Cao et al., 2010

Notes: R_Red, R_NIR, R_SWIR1 and R_SWIR2 demonstrate the reflectance of the Red band, Near Infrared band, Shortwave Infrared 1 band(1600 nm) and Shortwave Infrared 2 band (2100 nm), corresponding to the band 4, 8, 11 and 12 of sentinel-2A data, respectively.

).

PVIs and NPVIs were calculated using ArcGIS 10.5 software. The PVI value and f_PV measured value, and NPVI value and f_NPV measured value, were constructed linear regression models of PVI and NPVI relative to f_PV and f_NPV, respectively, and were used to evaluate the accuracy of the various indices. Considering the number of samples we have taken, we used the leave-one-out cross-validation (LOOCV) which was suitable for a small number of sample. In LOOCV, each sample was excluded in turn and the regression model was calculated with all the remnants samples and used to predict that sample. The benefit of LOOCV was its aptitude to detect outliers and its capability of providing nearly unbiased estimations of the prediction error [4,37]. The performance of these models was assessed by the coefficients of determination (R²), root mean square error of leave-one-out cross-validation (RMSECV) and regression significance (p):

$$R^2=\frac{{{\displaystyle \sum }}_{i=1}^n\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}{{{\displaystyle \sum }}_{i=1}^n{(x_i-\overline{x})}^2{{\displaystyle \sum }}_{i=1}^n{(y_i-\overline{y})}^2}$$

(1)

$$RMSECV=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(x_i-y_i\right)}^2}$$

(2)

where n shows the sample plots’ number, x_i is the measured value of the sample plot i, y_i stands for the estimated value of the sample plot i, $\overline{x}$ illustrates the mean value of the measured sample plots, $\overline{y}$ stipulates the mean value of the estimated sample plots.

2.3.2. Linear Unmixed Model

Guerschman [3] hypothesized that the NDVI and CAI could resolve the fractions of the PV, NPV and BS, when the NDVI and f_PV were linearly related (as were the CAI and f_NPV). This situation is reflected in the scatterplot of the NDVI and CAI, namely that the feature space forms a triangle; BS is situated on the right side of the triangle, with a high NDVI and intermediate CAI value; the NPV is located in the upper left corner of the triangle, showing a low NDVI and a high CAI value; the BS is seen in the lower left corner of the triangle, having low NDVI and CAI values; Cao [24] used the DFI (replacing CAI) to present the NPV in order to construct the NDVI-DFI linear unmixed model, and it is successful in estimating the fractions of the PV, NPV and BS. Therefore, we applied the GEMI-DFI linear unmixed model to assess the fractions of the PV, NPV and BS in the study area (Figure 4).

PV is expressed by the global environmental monitoring index (GEMI) and the NPV by the dead fuel index (DFI), The GEMI formula is calculated as Table 1, The DFI formula is the following in Table 2. The relative proportions of each fractional coverage for any Sentinel-2A image pixel were found by solving the equations:

$$G_S={\displaystyle \sum }\left[f_i{G}_i\right]=\left[f_{\mathrm{PV}}{G}_{\mathrm{PV}}+f_{\mathrm{NPV}}{G}_{\mathrm{NPV}}+f_{\mathrm{BS}}{G}_{\mathrm{BS}}\right]$$

(3)

$$D_S={\displaystyle \sum }\left[f_i{D}_i\right]=\left[f_{\mathrm{PV}}{D}_{\mathrm{PV}}+f_{\mathrm{NPV}}{D}_{\mathrm{NPV}}+f_{\mathrm{BS}}{D}_{\mathrm{BS}}\right]$$

(4)

$${\displaystyle \sum }{f}_i=\left[f_{\mathrm{PV}}+f_{\mathrm{NPV}}+f_{\mathrm{BS}}\right]=1$$

(5)

where G_S and D_S show the GEMI and DFI value in the given Sentinel-2A image pixel, the f_PV, f_NPV and f_BS illustrate the fractional coverage of the PV, NPV and BS; the G_PV, G_NPV and G_BS are the GEMI values of the end members, the D_PV, D_NPV and D_BS demonstrate the DFI values of the end members. It forces the values of f_PV, f_NPV and f_BS to sum to unity. If the sum does not get one, the pixel has a negative or higher than 1 value in at least one end member. When that occurred, the following correction was applied:

$$C_x=0(-0.2<C_x<0)$$

(6)

$$C_x=1(1<C_x<1.2)$$

(7)

$$C_y=C_y/\left(C_y+C_z\right)$$

(8)

$$C_z=C_z/\left(C_y+C_z\right)$$

(9)

where C_x is the value not within a specified range (C_x < −0.2 or C_x > 1.2), and C_y and C_z are the values of the other two end members. If the range of C_x is amounts from −0.2 to 0, C_x is 0; if the range of C_x is 1 to 1.2, C_x is 1; the condition is mentioned in the above two cases, we only calculated C_y and C_z.

2.3.3. Determination of the GEMI and DFI End Member Value

The choice of pure end members was the key to success of the PV and NPV model inversion. The GEMI-DFI linear unmixed model needed to become a pure end member of the PV, NPV and BS to calculate the corresponding proper values and the model needed be used to solve the fractional coverage of each end member. We employed the pixel purity index (PPI) method to determine the pure end member. Firstly, the Sentinel-2A images were subjected to the minimum noise fraction (MNF) so as to reduce the image dimensionality during different periods. The first 6 bands of each image period were selected for calculation and the number of iterations was set to 5000 in order to generate the PPI. Secondly, the GEMI and DFI were calculated for various period images. Finally, the pixels near the vertices of the triangular feature space were regarded as pure pixels.

2.3.4. Model Evaluation

In order to quantitatively evaluate the performance of the model’s estimation ability, the coefficients of determination (R²), root mean square error (RMSE) and mean error (ME) were applied.

$$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum }_{i=1}^n{\left(x_i-y_i\right)}^2}$$

(10)

$$ME=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left(x_i-y_i\right)$$

(11)

where n shows the sample plots’ number, x_i is the measured value of the sample plot i, y_i stands for the estimated value of the sample plot i.

3. Results

3.1. PVI and NPVI Index Optimization

In this study, six PV indices (NDVI, RVI, SAVI, MSAVI, RSR and GEMI) were selected to build linear regression models with f_PV (

Table 3. Parameters of linear regression analyses between PVIs and f_PV.

PVIs	Regression Equation	R2	RMSECV	p
NDVI	y = 1.4562x + 0.1199	0.51	0.0815	p < 0.05
RVI	y = 0.9331x + 0.7924	0.47	0.0817	p < 0.05
SAVI	y = 1.1202x − 0.1199	0.51	0.0795	p < 0.05
MSAVI	y = 0.9333x − 0.1407	0.47	0.0814	p < 0.05
RSR	y = 0.4316x − 0.2178	0.33	0.1283	p < 0.05
GEMI	y = 2.8495x − 0.0349	0.59	0.0752	p < 0.05

and Figure 5). Various PVIs exhibited different performance, with the R² ranging from 0.33 to 0.59 and the RMSECV ranging from 0.752 to 0.1283. The GEMI index showed the highest correlation, with the R² reaching 0.59 and RMSECV being 0.752 (p < 0.05). The RVI, MSAVI and RSR indices had low correlations (all R² lower than 0.50 and RMSECV higher than 0.8). The GEMI index was used as the PVI index to estimate the f_PV of the study area. In order to count the GEMI value, the GEMI index value had been normalized.

We selected six NPVI indices (NDI, NDTI, NDSAVI, DFI, STI and SWIR32) to construct linear regression models with the f_NPV (

Table 4. Parameters of linear regression analyses between NPVIs and f_NPV.

NPVIs	Regression Equation	R2	RMSECV	p
NDI	y = 0.9950x + 0.4855	0.02	0.2627	p < 0.05
NDTI	y = 3.4457x + 0.1076	0.39	0.2223	p < 0.05
NDSVI	y = 3.1160x + 0.9570	0.11	0.2606	p < 0.05
DFI	y = 0.0328x − 0.0422	0.45	0.2111	p < 0.05
STI	y = 1.4480x − 1.3296	0.43	0.2161	p < 0.05
SWIR32	y = −2.0018x + 2.0096	0.37	0.2272	p < 0.05

and Figure 6). Compared with the correlation between the PVI indices and f_PV, the correlation between the NPVIs and f_NPV was lower. This might be caused by similar spectral characteristics of the NPV and BS [24]. The DFI index showed the best performance with a R² of 0.45 and RMSECV of 0.2111(p < 0.05). The STI, NDTI, SWIR32, NDSVI and NDI indices were lower than the DFI, with the R² ranging from 0.02 to 0.43, and the RMSECV from 0.2611 to 0.2627. Therefore, the DFI index was selected to assess the f_NPV of the study area.

3.2. The Feasibility of the GEMI-DFI Model

We obtained the GEMI-DFI response space by calculating the GEMI and DFI values for seven Sentinel-2A images’ data (Figure 7). GEMI-DFI response space of the first six images was basically triangular, which was consistent with the theoretical conceptual mode. It showed that the GEMI-DFI linear unmixed model could be utilized to estimate the fractional coverage of the PV, NPV and BS. The last GEMI-DFI response space image did not seem to conform to the triangle shape. We selected images covering the growing season from spring to fall (April to November) and as vegetation leafed out and became active, GEMI values increased while DFI values declined. The inverse was seen with the onset of fall, as vegetation began to senesce and go dormant.

3.3. Evaluation of the f_PV and f_NPV Estimation Accuracy

Considering that field data collection occurred between 25–29 September in 2020, Sentinel 2A data collected on 2 October in 2020 were selected. We used the GEMI-DFI linear unmixed model (combined with the PPI) to obtain the spatial distribution of the f_PV, f_NPV and f_BS. The BS proportion in the study area was the largest and the PV and NPV were mainly distributed along the downstream river and the Tetima Lake. In order to clearly show the model result in estimating the f_PV and f_NPV, we selected the areas with a vegetation distribution evenly for display. The four regions were: a (Dashkol), b (Chiwinkol), c (Bozkol) and d (Tetima Lake). The f_NPV of the four regions was greater than the f_PV, indicating that after October, the vegetation entered the end of the growing season and the increase in the NPV (during this period) caused an increase in the proportion of the f_NPV (Figure 8) in turn.

We could conclude from Figure 9 that the estimated f_PV has the best correlation with the measured f_PV, with a R² of 0.69 and RMSE of 0.07 (p < 0.05). The estimated f_NPV and the measured f_NPV have a lower correlation with a R² of 0.58 and RMSE of 0.17 (p < 0.05). For BS, the measured (and estimated) values possess the lowest correlation, with a R² of 0.43 and a RMSE of 0.17 (p < 0.05). For the PV, the fitted line is situated below the reference line (1:1 line) and the ME value amounts to −2.67%. The data mentioned above show that overall the estimated f_PV value is lower than the measured f_PV value, this is due to the fact that the acquisition time of the Sentinel-2A image data is later than the one for the field measured data, resulting in lower f_PV estimates. Concerning the f_NPV, the value of ME is 3.14%, this means that the estimated f_PV value is higher than the measured f_PV value. In general, the estimation accuracy of the f_NPV and f_BS is lower than the f_PV, this reflects the fact that the PV could be resolved in an easier way than the NPV. Both the NPV and BS have similar spectral reflectance characteristics, resulting in a greater probability of wrong classification between the NPV and BS.

3.4. Seasonal Variation of the f_PV and f_NPV

In this study, the b region was chosen as the representative area for seasonal variation. This was located in the Chiwinkol wetland, which was less disturbed by human activities and could have better indicated the natural alternation of the PV and NPV’s seasonal variation. We selected Sentinel-2A image data from seven periods and used the GEMI-DFI linear unmixed model to estimate the f_PV, f_NPV and f_BS during different periods (Figure 10). We counted mean values of the f_PV, f_NPV and f_BS in the b region (Figure 11). By analyzing the changes of the f_PV, f_NPV and f_BS values, it was found that the seasonal variation of the f_PV and f_NPV was estimated by the GEMI-DFI model in conformity with the characteristics of the vegetation phenology. In April, most of the vegetation not yet broken dormancy and as such, there was a large NPV amount, the GEMI value was low, the DFI value was high, the f_NPV was 0.93 and the f_PV measured 0.01. In May, as the vegetation had emerged and begun actively growing, the GEMI value increased, the DFI value decreased, the f_PV rose to 0.17 and the f_NPV declined to 0.69. As the growing season progressed, GEMI increased, peaking in August with DFI following an inverse trend, and the DFI value fell to the lowest point. Similarly, f_PV reached maximum value, which was 0.44 and f_NPV its lowest value of 0.45. In September, the vegetation began to yellow and GEMI fell and DFI rose. By the beginning of October, most of the vegetation had already withered, the f_NPV occupied the dominant position again (at 0.81) and the f_PV decreased further to 0.17. The f_BS value remained in a relatively stable position throughout the whole year.

4. Discussion

This study explored the performance of several Sentinel-2A based PVIs by constructing linear regression models in the lower reaches of the Tarim River. The GEMI index has the best correlation with the f_PV, which is consistent with the conclusion of Liu [38] on the vegetation information monitoring this area. The GEMI index shows greater advantages in detecting the low coverage vegetation information. In the regression analysis between the NPVI index and f_NPV, the correlation between the DFI index and f_NPV is better than for those of the NDI, NDTI, NDSVI, STI and SWIR32. Among them, the NDI and NDSVI indices mainly use Band 11 (SWIR1: 1565~1655 nm), combined with a visible light band (NDI for the NIR band, NDSVI for the Red band) and without combining Band 12 (SWIR: 2100~2280 nm). Previous research has demonstrated that the cellulose absorption signature (amounting to around 2100 nm in the SWIR region) has a strong discriminatory ability between the NPV, PV and BS [8]; so Band 12 is a good choice to distinguish the PV, NPV and BS. Compared with the DFI index, STI and SWIR32 simply perform ratio calculations. The DFI index uses more bands for the calculation, including the Red, NIR SWIR1 and SWIR2. It might reduce the impact of the BS background to some extent and it might further improve NPV detection [4]. Ji [39] similarly concluded that the red-edge and NIR bands of the Sentinel-2 data are effective in improving the accuracy of the f_NPV estimates. The manner in which the red-edge band of the Sentinel-2 data could be utilized in order to estimate the vegetation information accurately still need further exploration.

The study uses the PPI method to extract the pure end members of the image data, which eliminates the discrepancies between the image data and the measured data, ensuring that both have the same spatial scale, which is widely employed to select the pure end of the spectrum [14,37]. We were able to estimate the f_PV and f_NPV of the study area in different periods by means of the GEMI-DFI model, revealing the changes in the photosynthetic and non-photosynthetic vegetation at different growth stages. The NPV proportion is greater at the beginning and at the end of the growing season, while the proportion of the PV peaks in the middle of the vegetation growing season. This is consistent with the findings of Wang [40] estimating the f_PV and f_NPV in the typical grasslands of Xilingol, indicating that the GEMI-DFI model is feasible to estimate the f_PV and f_NPV in the lower reaches of the Tarim River. The accuracy of the f_NPV estimation (R² = 0.58) is slightly lower than for the f_PV estimation (R² = 0.69). As the NPV and BS have similar spectral reflection characteristics in the visible band [3,41], the NPV estimation is more susceptible to the influence of the BS background, which could lead to a false distinction between the NPV and BS. Both the GEMI and DFI values are influenced by a variety of factors, such as the vegetation type, vegetation structure, etc. [17,24]. Therefore, more consideration should be given to the influencing factors in future studies on the PV and NPV.

The overall vegetation coverage in the study area is low, as vegetation is mostly distributed along the watercourses. Due to this, higher spatial resolution imagery is required to obtain more detailed information on the type of the vegetation [42]. Based on this, the Sentinel-2A data with a resolution of 10 m were used, which might reduce the spatial heterogeneity of the mixed pixels and improve the estimation accuracy by a better acquisition of pure pixels. There is still some uncertainty noticeable in the estimation of the PV, NPV and BS using multispectral data. In fact, the vegetation and other surrounding components form a complex system [37] and it endures stages of greening, heading, flowering, maturing and yellowing [43]. The above-mentioned conditions affect the acquisition of the estimated GEMI and DFI values, thereby impacting the model’s accuracy to estimate the f_PV, f_NPV and f_BS. In addition, uncertain factors also exist in the field data collection. The moment in which the photo was taken does not match the time of the satellite transit; the visual interpretation of the PV, NPV and BS classification and the accuracy of the geometric alignment will inevitably be influenced by subjective factors, resulting in errors in the f_PV, f_NPV and f_BS estimates [18]. This study relies on the coverage obtained from field sampled data as the verification data. Taking into account the diversity of vegetation types and structures, the spectral characteristics of vegetation are different. We should combine spectral data to accurately estimate the changing trends of different types of vegetation. The use of drones is also an effective measure to improve the estimation of vegetation coverage. This method can obtain a large area of vegetation coverage and reduce the error of its estimation and should be taken into account in future research work.

5. Conclusions

In this study, we constructed the GEMI-DFI linear unmixed model based on the Sentinel-2A image data so as to estimate the fractional coverage of the PV, NPV and BS in the lower reaches of the Tarim River combined with the field measured data for an accuracy evaluation. The main conclusions are as follows:

We established a linear regression model for the PVIs and f_PV, and NPVIs and f_NPV. The study found that the GEMI have a significantly linear correlation with the f_PV, R² is 0.59 and the DFI has a significantly linear correlation with the f_NPV, R² measuring 0.45. We used GEMI and DFI indices to construct linear unmixed model, the response space shown as triangle, which conformed the basic assumption of the linear unmixed model.

The GEMI-DFI linear unmixed models could effectively estimate the f_PV and f_NPV, but the accuracy of estimation f_PV is higher than that of f_NPV. How to improve the accuracy of estimation of f_NPV is the focus of future work.

Considering the number of sampling data, we need to collect more vegetation coverage data in future work. We can use drones as a platform for obtaining vegetation coverage to achieve a wide range of coverage, thereby improving the accuracy of f_PV, f_NPV and f_BS estimation.