What Is the Effect of Quantitative Inversion of Photosynthetic Pigment Content in Populus euphratica Oliv. Individual Tree Canopy Based on Multispectral UAV Images?

Abstract

It is highly necessary to apply unmanned aerial vehicle (UAV) remote sensing technology to forest health assessment. To prove the feasibility of quantitative inversion of photosynthetic pigment content (PPC) in Populus euphratica Oliv. individual tree canopy (PeITC) by using multispectral UAV images, in this study, Parrot Sequoia+ multispectral UAV system was manipulated to collect the images of Populus euphratica (Populus euphratica Oliv.) sample plots in Daliyabuyi Oasis from 2019 to 2020, and the canopy PPCs of five Populus euphratica sample trees per plot were determined in six plots. The Populus euphratica crown regions were extracted by grey wolf optimizer-OTSU (GWO-OTSU) multithreshold segmentation algorithm from the normalized difference vegetation index (NDVI) images of Populus euphratica sample plots obtained after preprocessing, and the PeITCs were segmented by multiresolution segmentation algorithm. The mean values of 27 spectral indices in the PeITCs were calculated in each plot, and the optimal model was constructed for quantitative estimation of the PPCs in the PeITCs, then the inversion results were compared and verified based on GF-6 and ZY1-02D satellite imageries respectively. The results were as follows. (1) The average value of canopy chlorophyll content (Chl) was 2.007 mg/g, the mean value of canopy carotenoid content (Car) was 0.703 mg/g. The coefficient of variation (C.V) of both were basically the same and they were both of strong variability. The measured PPCs of the PeITCs in Daliyabuyi Oasis was generally low. The average contents of chlorophyll and carotenoid in PeITC in June were more than twice those in August, while the mean ratio between them was significantly lower in June than in August. The measured PPCs had no obvious spatial distribution law. However, that could prove the rationality of sample selection in this study. (2) NDVI had the best effect of highlighting vegetation among all quadrats in the study area. Based on the GWO-OTSU multithreshold segmentation method, the canopy area of Populus euphratica could be quickly and effectively extracted from the quadrat NDVI map. The best segmentation effect of PeITCs was obtained based on a multiresolution segmentation method when the segmentation scale was 120, the shape index was 0.7, and the compactness index was 0.5. Compared with manual vectorization method of visual interpretation, the root mean square error (RMSE) and Pearson correlation coefficient (R) values of the mean NDVI values in PeITCs obtained by these two methods were 0.038 and 0.951. (3) Only 12 of the 27 spectral indices were significantly correlated with Chl and Car at the significance level of 0.02. Characteristics of the calibration set and validation set were basically consistent with those of the entire set. The classification and regression tree-decision tree (CART-DT) model performed best in the estimation of the PPCs in the PeITCs, in which, when estimating the Car, the calibration coefficient of determination (R²_C) was 0.843, the calibration root mean square error (RMSE_C) was 0.084, the calibration residual prediction deviation (RPD_C) was 2.525, the validation coefficient of determination (R²_V) was 0.670, the validation root mean square error (RMSE_V) was 0.251, the validation residual prediction deviation (RPD_V) was 1.741. (4) Qualitative comparison of spectral reflectance and NDVI values between GF-6 multispectral imagery and Parrot Sequoia+ multispectral image on the 172 PeITCs can show the reliability of Parrot Sequoia+ multispectral image. The comparison results of five PeITCs relative health degree judged by field vision judgment, measured SPAD value, predicted value of Chl (Chl_pre), the red edge value calculated by ZY1-02D (ZY1-02D_red edge) and the Carotenoid Reflection Index 2 (CRI2) value calculated by ZY1-02D (ZY1-02D_CRI2) can further prove the scientificity of inversion results to a certain extent. These results indicate that multispectral UAV images can be applied for quantitative inversion of PPC in PeITC, which could provide an indicator for the construction of a Populus euphratica individual tree health evaluation indicator system based on UAV remote sensing technology in the next step.

1. Introduction

The increasing human demand for forests has brought unprecedented attention to forest health issues [1], and forest health assessment theories and technical methods have become one of the most important and urgent research topics in the field of environmental and ecosystem sciences in the 21st century [2]. At the same time, with the development of remote sensing technology, it is feasible and necessary to apply remote sensing technology to forest health assessment [3,4,5], which can effectively make up for the shortage of time-consuming and labor-intensive traditional survey methods. The application of remote sensing technology has also always been the hot spot and frontier in the field of forest health assessment [6,7,8,9].

Before the advent of unmanned aerial vehicle (UAV) technology, satellites and manned aircraft were the main remote sensing platforms [10]. The remote sensing imageries acquired based on these two remote sensing platforms were limited by the relatively low spatial resolution [11]. It could not be effectively applied to the study of individual tree scale health evaluation [12]. This had led to a long period of slow progress in this research area. Fortunately, the UAV platform had managed to break through this limitation with its centimeter-level spatial resolution [10]. It made it possible to evaluate the health of individual trees based on UAV remote sensing technology [13]. In recent years, the emergence of consumer multispectral UAVs has provided a new perspective for real-time insight into forest health. Compared with ordinary RGB cameras, multispectral cameras can provide more vegetation spectral information [14]. Compared to hyperspectral cameras, multispectral cameras cannot compete in terms of the fineness of spectral information. However, at present, the price of hyperspectral cameras is often prohibitive, and the complexity [15] of hyperspectral information has made the more simple and practical consumer multispectral UAVs widely used with their affordable price.

Czapski et al. preliminarily analyzed the forest health status based on UAV multispectral images [16]. Dash et al. tested the sensitivity of detecting the degree of herbicide stress on artificial radiata pine from time-series multispectral images acquired from UAV and satellite platforms, respectively, by means of controlled experiments. Their findings show that multispectral images collected from drones can successfully identify the level of physiological stress in mature plantation trees even in the early stages of tree stress [11]. It is also considered that multispectral UAV images are a powerful supplement for monitoring forest health [17]. Guerra-Hernández et al. developed a robust assessment method of alder forest health based on multispectral UAV images. The method quickly and effectively distinguishes areas affected by diseases from areas not affected by diseases, so as to determine key protection hotspots [18]. Hao et al. used consumer multispectral UAVs to evaluate the height and density of a young Chinese fir forest. Compared with the field survey results, at the stand level, the estimated coefficient of determination (R²) of tree height was 0.95, the root mean square error (RMSE) was 0.12 m, the estimated R² of tree density was 0.99, and RMSE was 48 trees per hectare. The results show that the consumer multispectral UAV can successfully monitor forest parameters. It has great potential to replace field investigation and forest health assessment [19]. Fraser and Congalton evaluated the ability of unmanned aerial systems (UAS) multispectral images and high-resolution aerial NAIP images to distinguish healthy, stressed and degraded individual trees in complex mixed forest. When using the random forest classifier, the classification accuracy of UAS images is 71.19% and that of aerial images is 70.62%. They think that further improving the accurate calibration of UAS multispectral images and improving the image segmentation method will further improve the accuracy of extracting forest health information from UAS multispectral images [13]. Abdallanejad and Panagiotidis used UAS multispectral images to classify tree species and assessed the health of coniferous and broad-leaved mixed forest. The aim was to develop a new workflow to improve the accuracy of tree species classification and detection of healthy, unhealthy and dead trees caused by bark beetles. The results emphasized the effectiveness of the proposed tree classification method, with the overall accuracy rate of 81.18% (Kappa: 0.70) and the accuracy rate of health assessment of 84.71% (Kappa: 0.66) [20]. Minařík and Langhammer proposed a new method to evaluate forest disturbance dynamics by combining multispectral imaging camera and UAV photogrammetry technology. The results showed that multispectral UAV images can detect tree stress caused by bark beetles and can correctly detect different forest disturbance categories under complex category mixing conditions [21]. Honkavaara et al. used multitemporal hyperspectral and multispectral UAV images to detect the early infestation of bark beetles on Norwegian spruce. The deterioration of tree health and the development of spectral symptoms were explored through the time series of hyperspectral UAV images. Then, based on a multispectral UAV image training machine learning model, spruce health status was classified into three grades. The preliminary results showed that the application of multispectral UAV in tree health assessment is promising [22].

To sum up, at present, most research on forest health assessment based on multispectral UAV images is to determine the health level by qualitative classification. Few studies have been done to evaluate forest health by quantitative inversion of forest canopy biochemical parameters. Previously, Kopačková-Strnadová et al. were the first to use the Parrot Sequoia+ multispectral UAV system to conduct a study on the canopy extent, top, height extraction and canopy photosynthetic pigment content (PPC) estimation of Norway spruce individual tree. The canopy top, height, and extent of Norway spruce individual trees were extracted by the canopy height model and watershed segmentation algorithm. The linear relationship between the measured canopy PPC and normalized difference vegetation index (NDVI) and red edge NDVI was then tested respectively. The effects of the needle age and the illumination conditions on the linear model were discussed. The results showed that NDVI and red edge NDVI extracted by a multipectral UAV system could be used to estimate the PPC in a Norway spruce’s individual tree canopy at a semiquantitative level. Furthermore, the needle age had a major influence on the estimation results, while the illumination conditions had little effect [23].

It is well known that the health degree of a forest canopy, as the part that can be effectively observed by remote sensing science, is a very important criterion layer factor in the forest health evaluation indicator system, and the canopy PPC is a key indicator layer factor to characterize the strength of photosynthesis and the health degree of forest canopy [24]. Gupta and Pandey developed a systematic method to evaluate forest health by using Sentinel 2A images involving inversion of chlorophyll content in the forest canopy by using various spectral indices. Through field investigation and verification, the correlation between canopy chlorophyll content (Chl) and forest health and the sensitivity of Chl to forest stress factors were proven [25]. Lou et al.’s quantitative estimation of swamp vegetation was based on multisource remote sensing data. A new method for obtaining multiscale sample data of chlorophyll content of swamp vegetation canopy based on UAV multispectral images was proposed. The application performance of GF-1 WFV, Landsat-8 OLI and Sentinel-2 MSI satellite remote sensing images in quantitative inversion of chlorophyll content in swamp vegetation canopy was evaluated by random forest regression algorithm [26]. Sun et al. believe that accurate inversion of Chl is very important for effective monitoring of forest productivity and environmental stress. Therefore, an integrated aggregation index (CI) method was proposed to retrieve forest Chl from a MERIS data set by empirical regression and random forest regression [27]. Gupta and Pandey used the PROSAIL radiation transfer model inversion (RTM) and an artificial neural network (ANN) to calculate the temporal and spatial heterogeneity of chlorophyll content in a subtropical forest canopy based on Landsat-8 and Sentinel 2A imageries. The results showed that from 1997 to 2017, the sharp decline of chlorophyll concentration caused various photosynthetic vulnerabilities in the forest ecosystem, leading to forest degradation and a significant decline in forest health [28]. Ali et al. used the radiation transfer model to generate Chl map of time series through RapidEye and Sentinel-2 (2011–2018) image sets. It was proven that the chlorophyll content of the canopy retrieved from time series remote sensing data can be used as an effective index to detect bark beetle infection in Picea abies [29].

From the above research, it can be concluded that quantitative inversion research of PPC in forest canopy is basically based on multispectral satellite images. However, there is little research based on multispectral UAV images. Moreover, no matter whether it is multispectral UAV or canopy PPC, the research on forestry is still immature (please refer to the Supplementary Materials), and there are still various problems that need to be further solved. Therefore, how to retrieve the PPC of forest canopy from multispectral UAV images becomes a quantitative remote sensing problem worthy of discussion at present [23].

In addition, under the background of global warming, natural oases in desert and semidesert areas are facing increasing environmental stress, and oasis ecological security is facing a severe test. Populus euphratica Oliv. (Populus euphratica or Euphrates poplar) is the only forest-forming tree species in desert and semidesert areas [30]. It is a very precious ancient and famous tree resource, and it is a key protected tree, as established by the United Nations Food and Agriculture Organization (FAO) [31]. At the same time, Populus euphratica forests play a pioneering role in preventing wind and fixing sand [32,33], and their health plays a key role in the ecological security of the distribution area. Therefore, it is very necessary to carry out the related research of Populus euphratica health evaluation. Li et al. believe that quantifying the phenological changes of Populus euphratica caused by climate change is very important for desert ecosystems. Therefore, they used the 18-year time series data of LAI (2000–2017) of the global surface satellite (GLASS) in the upper reaches of the Tarim River. The phenology of Populus euphratica and its response to climate change were studied. Finally, it was found that runoff was the more critical factor to control the phenology of Populus euphratica in this area [34]. Huang et al. based on MODIS NDVI time series data, using wavelet transform and discriminant analysis method, evaluated the damage degree of Apocheima cinerarius to the health of Populus euphratica forest. The results show that the prediction model achieved 91.7% and 94.37% accuracy of severity classification in detecting disease outbreak time. It is a rapid, accurate and practical method for detecting pest stress. It can be used to monitor and control the spring geometrid on desert Populus euphratica trees, which is of great significance to protect the desert ecological environment [35].

Populus euphratica forest, as a special riparian forest in arid desert area, is sparsely distributed. Therefore, it is more suitable for health monitoring and evaluation research from single tree scale, and that’s exactly what multispectral drones can do. In particular, no research on Populus euphratica has been carried out in this field. Therefore, in this study, a Populus euphratica tree was taken as the research object and the effectiveness of the inversion of PPC in Populus euphratica individual tree canopy (PeITC) using UAV multispectral images was taken as the core question. This study was dedicated to demonstrating the feasibility of multispectral UAV images for quantitative inversion of PPC in PeITC. The specific research process is shown in Figure 1.

In this paper, the following topics were studied. (1) The temporal and spatial distribution characteristics of PPC in PeITC in Daliyabuyi Oasis were explored based on the measured data of PPC in PeITC in August 2019 and June 2020. (2) The applicability of extracting PeITCs by grey wolf optimizer-OTSU (GWO-OTSU) multithreshold and multiresolution segmentation algorithm was tested based on Parrot Sequoia+ multispectral image data in August 2019 and June 2020. (3) The best quantitative estimation model was found between the measured PPC of PeITCs and the mean values of 27 spectral indices. (4) The reliability of the Parrot Sequoia+ multispectral images and the scientific validity of the inversion results obtained by the optimal quantitative estimation model were verified using GF-6 multispectral imageries and ZY1-02D hyperspectral imageries. This study can directly provide theoretical basis and methodological support for the health assessment of an individual Populus euphratica tree based on UAV remote sensing technology, thus contributing to further improve the efficiency of forestry health management and filling the relevant research gaps.

2. Materials and Methods

2.1. Study Area

Daliyabuyi Oasis is located in Yutian County, Xinjiang, China. As shown in Figure 2, it extends 250 km into the hinterland of the Taklamakan Desert, the second largest flowing desert in the world. It is a rare large-scale natural oasis in the desert hinterland [36]. Because of its isolation and inaccessibility, the oasis has been preserved in its original state [37] and has been called “Paradise” by archaeologists [38]. Daliyabuyi Oasis is 80 km long and 7–15 km wide, with an existing area of about 324 km² [39]. The area of natural Populus euphratica forest is about 31.3 km² [40]. The topography slopes from southwest to northeast, with elevations ranging from 1161–1212 m. The geomorphology is characterized by island dunes cut by rivers and scrub dunes, both of which form an irregular mosaic within the fan-shaped oasis [41]. The oasis is in a warm temperate extreme arid zone. It has a continental desert climate. The main climatic features are four distinct seasons, abundant light and heat, large temperature difference between day and night, very little precipitation, great evaporation, and disaster weather such as wind, sand and dust in spring and summer [42]. The natural vegetation is mainly fed by the reticulated river system formed by the lower reaches of the Keriya River and the shallow groundwater that seeps into the ground. Common constructive species are Populus euphratica, Tamarix chinensis (Tamarix chinensis Lour.) and Phragmites communis (Phragmites australis (Cav.) Trin. ex Steud.) [43]. The soil is mainly sandy soil [44]. At present, there are various indications that Daliyabuyi Oasis is threatened by severe desertification, the oasis ecosystem is in a highly unstable state, which is developing towards degradation [45,46].

2.2. Data

2.2.1. Multispectral Drone Images

In this study, a Parrot Bluegrass quadrotor UAV equipped with a Parrot Sequoia+ five-channel multispectral sensor and a light sensor (shown in Figure 2g) was used to acquire multispectral UAV images of Populus euphratica quadrats. Parrot Sequoia+ multispectral sensors had been radiometrically calibrated prior to use and subsequently operated without radiometric calibration plates [47]. The UAV flight altitude was set to 50 m in 2019 and 30 m in 2020. Front overlap and side overlap were both 80%, and drone speed was normal. It was ensured that the UAV system could be fully warmed up before the flight at each sample site [48]. Before executing the flight plan, five typical Populus euphratica single trees with different growth conditions were selected visually in each quadrat, and they were guaranteed to be evenly distributed in the quadrat. Then five rigid-colored plates in white, black, yellow, red, and blue were marked around them respectively to ensure that the five marked Populus euphratica single trees could be identified and distinguished in the drone images. Afterwards, the Parrot UAV system was allowed to automatically execute the flight plan to acquire multispectral and RGB images of the sample plot. Finally, UAV images of Populus euphratica quadrats were successively obtained according to this workflow. Here, a further clarification is needed. To reflect the variability of Populus euphratica growth in Daliyabuyi Oasis more comprehensively and make the estimation model of PPC in PeITC more universal, the Populus euphratica quadrats in this study were selected by comprehensively considering the data of groundwater monitoring wells, investigation experience over the years, distribution uniformity, and Populus euphratica growth in the field.

In this research, the acquired Parrot drone multispectral and RGB images were preprocessed using Pix4Dmapper 4.4.10. Pix4Dmapper is capable of fully automatic processing of UAV images. Without manual intervention, thousands of images can be quickly produced into professional and accurate 2D orthophoto map. The main operation steps include: (1) project creation, (2) image input, (3) processing template setup, (4) initial processing, (5) point cloud and mesh, (6) DSM, orthomosaic and index. For the multispectral and RGB UAV images, Ag Multispectral and Ag RGB were selected in this paper when setting the processing templates in the Pix4Dmapper, respectively. The green band, red band, red edge band, near-infrared band reflectance orthophoto maps and RGB orthographic image were obtained for each Populus euphratica quadrat, respectively. After that, the layer stacking and clipping operations were performed in ENVI 5.3. Ultimately, the multispectral/RGB orthographic images of Populus euphratica quadrats could be used for research. The multispectral UAV orthophotography images had four bands with a ground resolution of 0.031–0.057 m. The specifications of the UAV orthophoto datasets are shown in

Table 1. The specifications of the UAV orthophoto datasets.

Quadrat	Number of Shooting Spots	Spatial Resolution (m)	Columns	Rows
2019-A	544	0.056	5000	5000
2019-B	282	0.056	5300	3100
2019-C	104	0.057	2600	2900
2019-D	117	0.052	3400	2400
2019-E	97	0.054	3000	2500
2019-F	286	0.057	3800	4300
2019-G	276	0.055	4500	4200
2020-A	490	0.033	5000	3000
2020-B	169	0.031	1700	2700
2020-C	560	0.033	8000	7200

.

2.2.2. Measured PPC Data

Forty-six labeled Populus euphratica sample trees were evenly distributed in the oasis, covering Populus euphratica with different health levels. This can be seen intuitively from Figure 5. The sampling period was from 16 August 2019 to 26 August 2019 (25 sample trees) and from 20 June 2020 to 27 June 2020 (21 sample trees). In both periods, the phenology of Populus euphratica in the study area was at the fruit ripening stage [49]. With reference to the sampling schemes of Niu et al. [50] and Wang et al. [51,52], at the time of field sampling in 2020, except for setting three Populus euphratica quadrats same as in 2019, a Populus euphratica sample tree which can represent the overall health status of Populus euphratica in each area was also selected near each groundwater level monitoring wells in Daliyabuyi Oasis for leaf collection to address the issue of sampling representativeness as much as possible. First, five leaves of different shapes were cut from four directions and the top of each sample tree using high branch shears [53]. Next, the SPAD value of each leaf was measured and recorded five times repeatedly using a SPAD-502 PLUS chlorophyll meter. Later, the five leaves were loaded together into a prepared aluminum case wrapped in black shell and labeled accordingly. Then, the aluminum case was placed in a cooler box filled with ice packs to keep the temperature down for the day. Upon returning to the field station at the end of each day’s trip, the aluminum cases in the cooler box were immediately put into the cryogenic refrigerator equipped for the field station. The above steps were repeated for each sample tree, and a total of 230 leaves were collected. After the field sampling program was completed, it was kept in low temperature and protected from light on the way back to the laboratory. When it was brought back to the laboratory, Chl and canopy carotenoid content (Car) were immediately measured using Agilent Cary 60 UV-Vis spectrophotometer with reference to NY/T 3082-2017 national standard (unit: mg/g). The mean value of five leaves was taken as the content of chlorophyll and carotenoid in the PeITC. The calculation equations of PPC used in this study are shown as Equations (1)–(4):

$$Chl-a=(12.72\times A_1-2.59\times A_2)\times \upsilon /(1000\times m)$$

(1)

$$Chl-b=(22.88\times A_2-4.67\times A_1)\times \upsilon /(1000\times m)$$

(2)

$$Chl=(Chl-a)+(Chl-b)=(8.05\times A_1+20.29\times A_2)\times \upsilon /(1000\times m)$$

(3)

$$Car=[1000\times A_3-3.27\times (Chl-a)-104\times (Chl-b)]/229\times \upsilon /(1000\times m)$$

(4)

where $A_1$ is the absorbance value of the test solution at 663 nm, $A_2$ is the absorbance value of the test solution at 645 nm, $A_3$ is the absorbance value of the test solution at 470 nm, $\upsilon$ indicates the volume of the test solution (mL), $m$ indicates the mass of the sample (g) [54,55,56]. Three replicate experiments were set up for each leaf, and the calculated results were all kept in three valid digits, and the final results were taken as the arithmetic mean of the three replicate experiments.

2.2.3. GF-6 PMS Multispectral Imagery

The GF-6 satellite runs in a solar synchronous return orbit at an altitude of about 645 km, with a descending node at local time of 10:30 a.m. It is equipped with a high-resolution camera and a wide-width camera [57]. In this paper, two multispectral imageries of Daliyabuyi Oasis, taken by high-resolution camera on August 22, 2019, were used to carry out the study, which contained a 2 m-resolution panchromatic band and an 8 m-resolution blue, green, red, and near-infrared bands [58]. GF-6 PMS imageries were preprocessed in ENVI 5.3. The pretreatment process mainly included radiometric calibration, FLAASH atmospheric correction and orthorectification of multispectral imageries, and then radiometric calibration and orthorectification of panchromatic imageries. Among them, the orthorectified correction was done with ASTER GDEM V2 30 m elevation data. Then the panchromatic and multispectral imageries were fused using the Gram–Schmidt fusion method (after testing, the Gram–Schmidt fusion method worked best). Eventually, the 2 m spatial resolution GF-6 multispectral imageries were obtained for subsequent research in this study.

2.2.4. ZY1-02D VNIC/AHSI Hyperspectral Imagery

ZY1-02D satellite was launched on 12 September 2019, carrying a visible/near-infrared camera and a hyperspectral camera. The visible/near-infrared camera can acquire a 2.5 m spatial resolution panchromatic band and eight 10 m spatial resolution multispectral bands including the red edge band, with a width of 115 km [59]. The hyperspectral camera can acquire 166 hyperspectral bands with 30 m spatial resolution (in the range of 400–2500 nm). It contains 76 visible/near-infrared bands with 10 nm spectral resolution and 90 short-wave infrared bands with 20 nm spectral resolution. Its width reaches 60 km [60]. ZY1-02D satellite imageries used in this study were taken in Daliyabuyi Oasis on 15 June 2020. To obtain the ZY1-02D hyperspectral imagery with 2.5 m spatial resolution, the panchromatic imagery acquired by the visible/near-infrared camera and the visible/near-infrared spectral band imagery acquired by the hyperspectral camera were utilized in this study. They were preprocessed in ENVI 5.3, and the pretreatment process was the same as that of GF-6 imageries. After that, the Gram–Schmidt fusion method was also applied to fuse the two imageries (After testing, the Gram–Schmidt fusion method was still the best). Lastly, the 2.5 m spatial resolution ZY1-02D hyperspectral imagery used in the next step was generated successfully.

2.3. Methods

2.3.1. GWO-OTSU Multithreshold Segmentation Algorithm

The OTSU threshold segmentation algorithm was proposed by Otsu in 1979 [61], which is an adaptive threshold determination method. The criteria for picking the segmentation threshold are that the interclass variance of the image is maximized or the intra-class variance is minimized. The OTSU threshold segmentation method can be extended from single threshold to multithreshold segmentation. Multithreshold segmentation adopts different thresholds to segment the image into different regions or targets [62]. Applying intelligent algorithm to multithreshold search can greatly speed up the algorithm [63].

Assume that the image size is $M\times N$, the range of image gray levels is $[0,L-1]$, $n_i$ is the pixel point number of image gray level $i$, the occurrence probability of gray level $i$ is $p_i=n_i/(M\times N)$. Another assumption is that there are $n-1$ thresholds $T_1,T_2,\dots ,T_{n-1}$ to classify the image into $n$ categories, which is $C_0=\{0,1,\dots ,T_1\},\dots ,C_n=\{T_{n-1}+1,T_{n-1}+2,\dots ,L-1\}$, the occurrence probability of each category is denoted as $P_0,P_1,\dots P_{n-1}$, the variance is ${\delta }_0^2,{\delta }_1^2,\dots ,{\delta }_{n-1}^2$, the mean value is $u_0,u_1,\dots ,u_{n-1}$. Then there are:

$$P_k={\displaystyle \sum _{i=T_k}^{T_{k+1}-1}{p}_i}$$

(5)

$$u_k=\frac{1}{P_k}{P}_k{\displaystyle \sum _{i=T_k}^{T_{k+1}-1}i\times p_i}$$

(6)

$${\delta }_k^2={\displaystyle \sum _{i=T_k}^{T_{k+1}-1}{(i-u_k)}^2\times \frac{p_i}{P_k}}$$

(7)

where, $k=0,1,\dots ,n-1$, $T_0=0$, $T_n=L$. Then the interclass variance of the image is expressed as:

$${\delta }_b^2={\displaystyle \sum _{i=0}^{n-1}{p}_i\times {\delta }_i^2}$$

(8)

The multilevel optimal segmentation threshold is:

$$\{T_1^{\ast },T_2^{\ast },\dots ,T_{n-1}^{\ast }\}=\arg {\max }_{(l\le T\le L)}\left\{{\delta }_b^2\right\}$$

(9)

The GWO algorithm was proposed by Mirjalili et al. in 2014 [64] as a novel population intelligence optimization algorithm. The GWO algorithm achieves the optimization goal by simulating the predation behavior of grey wolves and based on the cooperative mechanism of wolves. The GWO algorithm has the characteristics of a simple structure, few parameters to be adjusted and easy implementation. There are adaptive convergence factors and information feedback mechanisms. It can achieve a balance between local optimization and global search. As a result, it has good performance in terms of solution accuracy and convergence speed for the problem.

From the above principle of the OTSU multithreshold segmentation method, to get the final threshold value, it is necessary to look for the threshold value that makes the maximum value of interclass variance so the optimal threshold value can be obtained by using GWO algorithm for threshold hunting. Accordingly, the optimized fitness function is:

$$fun\{T_1^{\ast },T_2^{\ast },\dots ,T_{n-1}^{\ast }\}=\arg {\max }_{(l\le T\le L)}\left\{{\delta }_b^2\right\}$$

(10)

In practice, the number of threshold segmentation is simply set. The searching boundary is 0 to 255 (because the pixel values of the image range from 0 to 255), and the corresponding parameters of GWO algorithm can be set.

2.3.2. Multiresolution Segmentation Algorithm

Multiresolution segmentation is one of the most common segmentation methods in eCognition software. The principle of multiresolution segmentation is to set the segmentation scale to minimize the internal heterogeneity of the image object formed by the whole image segmentation through the selection of spectral and shape feature parameters [65].

$$h_{spectral}={\displaystyle \sum _c^{}{\omega }_c\times {\sigma }_c}$$

(11)

where $h_{spectral}$ is the spectral heterogeneity, ${\omega }_c$ is the layer weight, ${\sigma }_c$ is the layer standard deviation (SD), and $c$ is the amount of layers.

$$h_{shape}={\omega }_{smoothness}\times h_{smoothness}+{\omega }_{compactness}\times h_{compactness}$$

(12)

$$h_{smoothness}=n_{merge}\times \frac{l_{merge}}{b_{merge}}-(n_{obj1}\times \frac{l_{obj1}}{b_{obj1}}+n_{obj2}\times \frac{l_{obj2}}{b_{obj2}})$$

(13)

$$h_{compactness}=n_{merge}\times \frac{l_{merge}}{\sqrt{n_{merge}}}-(n_{obj1}\times \frac{l_{obj1}}{\sqrt{n_{obj1}}}+n_{obj2}\times \frac{l_{obj2}}{\sqrt{n_{obj2}}})$$

(14)

where $h_{shape}$ is shape heterogeneity, $h_{smoothness}$ and $h_{compactness}$ are smooth heterogeneity and compact heterogeneity respectively, ${\omega }_{smoothness}$ and ${\omega }_{compactness}$ represent the weight between the two indicators, and ${\omega }_{smoothness}+{\omega }_{compactness}=1$, $l$ indicates the side length of the image object, $b$ is the shortest side length of the image object, and $n$ is the area of the image object [66].

$$h={\omega }_{spectral}\times h_{spectral}+{\omega }_{shape}\times h_{shape}$$

(15)

$h$ is the integral heterogeneity, ${\omega }_{spectral}$ and ${\omega }_{shape}$ represent the weights between spectral and shape heterogeneity, and ${\omega }_{spectral}+{\omega }_{shape}=1$.

The workflow of multiresolution segmentation algorithm is roughly as follows: (1) setting segmentation scale parameters and determining the weights of spectral heterogeneity and shape heterogeneity according to the spectral, shape, texture, and other characteristics of each target in remote sensing images. (2) using the bottom-up segmentation method, selecting any pixel in the image as the initial growth seed of the first segmentation and calculating the initial heterogeneity parameters. (3) the newly generated image area is used as the image base to participate in the next segmentation, and the segmentation continues when $h<s^2$ ($s$ is the threshold of multiresolution segmentation), and it is completed when $h\ge s^2$ [67]. This cycle continues until all image elements in the entire image have been segmented into the corresponding image objects.

2.3.3. Statistical Regression Model

A total of seven different statistical regression models were tested in this study to maximize model estimation performance, as follows:

(1).

Classification and Regression Tree (CART)

The classification and regression tree-decision tree (CART-DT) regression model, as a classical machine learning model, introduces a binary cut to handle continuous data and is suitable for modeling complex data with multiple feature variables. It has the merits of simple extraction rules, high accuracy, and strong interpretability [68]. However, due to the poor stability of the algorithm in some cases, it is prone to overfitting [69]. This gives rise to the CART Ensembles and Bagger (CART-EB) regression model. CART-EB grows the decision trees in the ensemble using a bootstrap sample of the data, which reduces the effects of overfitting and improves generalization [70].

The main construction process of the CART-DT regression model includes decision tree generation and decision tree pruning. In this case, generating a CART regression tree is the process of recursively constructing a binary decision tree. The least square error principle is adopted to select features and generate the least square tree [71]. On the other hand, CART decision trees are usually based on the validation dataset using cost complexity pruning on the generated subtrees to obtain the optimal subtree with the minimum loss function. In the following, the algorithm for generating CART regression trees is presented in this study which is as below.

If $x$ and $y$ are input and output variables, respectively, and $y$ is a continuous variable, $n$ is the number of variables, the modeling data set $A$ is given as:

$$A=\{(x_1,y_1),(x_2,y_2),\cdots ,(x_n,y_n)\}$$

(16)

If the current cutting variable is assumed to be the $h$ th variable and the corresponding cutting feature value is $m$, two regions $\alpha$ can be divided and defined as:

$$\{\begin{array}{c}{\alpha }_1(h,m)=\left\{x|x^{(h)}\le m\right\}\\ {\alpha }_2(h,m)=\left\{x|x^{(h)}>m\right\}\end{array}$$

(17)

By Equations (16) and (17), the input space will be constantly partitioned into $T$ subregions ${\alpha }_1,{\alpha }_2,\cdots ,{\alpha }_T$, and each subregion ${\alpha }_t$ contains part of the sample data $I$ and the output values ${\beta }_t$. The expression $f(x)$ of the model currently is:

$$f(x)={\displaystyle \sum _{t=1}^T[{\beta }_t\times I(x\in {\alpha }_t)}]$$

(18)

It follows from the principle of least squares that the mean value of all outputs $y_i$ on subregion ${\alpha }_t$ is the optimal value ${\widehat{\beta }}_t$ of ${\beta }_t$. The optimal cut-off point and the predicted value are then determined by the criterion of minimizing the squared error ${\displaystyle \sum _{}^{}{[y_i-f(x_i)]}^2}$. The expressions of the formulas are:

$${\widehat{\beta }}_t=AVG(y_i|x_i\in {\alpha }_t)$$

(19)

$$\underset{h,m}{\min }\left[\underset{{\beta }_1}{\min }{\displaystyle \sum _{x_i\in {\alpha }_1(h,m)}^{}{(y_i-{\beta }_1)}^2}+\underset{{\beta }_2}{\min }{\displaystyle \sum _{x_i\in {\alpha }_2(h,m)}^{}{(y_i-{\beta }_2)}^2}\right]$$

(20)

By traversing all the input feature variables and their feature values, the current optimal cut-off point $(h,m)$ can be found. Then the current space is divided into two subregions according to the cut-off point, at this point if these two subregions can no longer be divided, the corresponding optimal output value can be obtained, expressed as:

$$\{\begin{array}{c}{\widehat{\beta }}_1=AVG[y_i|x_i\in {\alpha }_1(h,m)]\\ {\widehat{\beta }}_2=AVG[y_i|x_i\in {\alpha }_2(h,m)]\end{array}$$

(21)

According to the above steps, if the region is still divisible, the above steps are repeated for the region until it stops when the whole regression tree is completely undivided.

(2).

Gaussian Process Regression (GPR)

GPR is a nonparametric probability model that uses the prior distribution of Gaussian process to analyze data. The model hypothesis of GPR includes both noise (regression residuals) and Gaussian process prior, which is solved by Bayesian inference, as described in the literatures [72,73]. GPR, as a new machine learning algorithm, has a good adaptability to deal with small sample, high-dimensional nonlinear and random data, and it has the benefits of hyperparameter adaptive projection, output distribution probability derivation and strong generalization ability. In this study, the GPR model to estimate PPC in PeITC was realized by appropriately modifying the code disclosed by Habibullah et al. [74].

(3).

Others

In addition, TGBM, RF, MARS and PLSR regression models were also tested. The introduction of these regression models is not described here. For more details, please refer to [75].

Finally, the following needs to be delivered. CART-DT, CART-EB, TGBM, RF and MARS regression models were constructed using Salford Predictive Modeler v8.2 software. GPR model was completed in Matlab 2019b software. PLSR model was employed in Xlstat 2019 software. When using these models, the default parameter settings were maintained and no optimization of model parameters was performed.

2.3.4. Performance Evaluation of Regression Model

In this study, three indicators, R², RMSE, and residual prediction deviation (RPD), were applied to measure the performance of the regression model. R² demonstrates the robustness of the model, RMSE characterizes the precision of the model, and RPD reflects the predictive power of the model [75]. Besides, Pearson correlation coefficient (R) was utilized to prefer the model independent variables. In general, the better the regression model, the larger the |R|, the larger the R², the smaller the RMSE, and the larger the RPD. SD and coefficient of variation (C.V) were used as statistical indicators of data dispersion [76]. The formulas for calculating these indicators are as follows:

$$R=\frac{{\displaystyle \sum _{i=1}^n[(X_i-\overline{X})\times (Y_{obs,i}-{\overline{Y}}_{obs,i})]}}{(\sqrt{{\displaystyle \sum _{i=1}^n{(X_i-\overline{X})}^2}})\times (\sqrt{{\displaystyle \sum _{i=1}^n{(Y_{obs,i}-{\overline{Y}}_{obs,i})}^2}})}$$

(22)

$$SD=\sqrt{\frac{{\displaystyle \sum _{i=1}^n(Y_{obs,i}}-{\overline{Y}}_{obs,i}{)}^2}{n}}$$

(23)

$$\overline{X}=\frac{{\displaystyle \sum _{i=1}^n{X}_i}}{n},\overline{Y}=\frac{{\displaystyle \sum _{i=1}^n{Y}_{obs,i}}}{n}$$

(24)

$$RMSE=\sqrt{\frac{{\displaystyle \sum _{i=1}^n{(Y_{obs,i}-Y_{\mathrm{mod}el,i})}^2}}{n}}$$

(25)

$$R^2=1-\frac{{\displaystyle \sum _{i=1}^n{(Y_{obs,i}-Y_{\mathrm{mod}el,i})}^2}}{{\displaystyle \sum _{i=1}^n{(Y_{obs,i}-{\overline{Y}}_{obs,i})}^2}}$$

(26)

$$RPD=\frac{SD}{RMSE}$$

(27)

$$C.V=\frac{SD}{\overline{Y}}\times 100\%$$

(28)

where $i$ means the $i$ th PeITC, $n$ means the amount of PeITCs, $X_i$ means the spectral index value of the $i$ th PeITC, and $\overline{X}$ indicates the mean spectral index value of all PeITCs. $Y_{obs,i}$ denotes the measured PPC of the $i$ th PeITC, ${\overline{Y}}_{obs,i}$ denotes the mean value of the measured PPCs in all PeITCs, and $Y_{\mathrm{mod}el,i}$ denotes the estimated value of PPC in the $i$ th PeITC.

Figure 3 shows the empirical judgment criteria of R², RPD and C.V value respectively [77,78,79]. This criterion was adopted in the subsequent study of this paper to determine the regression model performance and data dispersion intervals.

3. Results and Discussion

3.1. Analysis of the Measured PPCs in the PeITCs

Photosynthetic pigments are important functional components for the perception, absorption, and utilization of solar energy by plants. Photosynthetic pigments are in a relative equilibrium state of continuous formation and decomposition in plants. Their content affects the photosynthetic capacity of plants [80]. Photosynthetic pigments include two major categories: chlorophyll, mainly chlorophyll a, b, c, d, and f. Chlorophyll a is the most abundant in terrestrial plants. Chlorophyll b is present in higher plants as an auxiliary photosynthetic pigment. The ratio of chlorophyll a to chlorophyll b in vegetation is usually around 2.5–4.0 [80], and the ratio of chlorophyll a to chlorophyll b decreases with decreasing photometric density. A low chlorophyll a to chlorophyll b ratio enhances the absorption of far-red light by plants [81]. While chlorophylls c, d, and f are only found in algae and cyanobacteria. Another group is carotenoids, which are water-resistant pigments with photosynthetic and photoprotective effects. Leaf heat dissipation capacity is closely related to carotenoid content and its components. Carotenoid content increases in leaves of plants growing under strong light conditions [82]. Carotenoids are mainly divided into two types: carotenoids and lutein. Studies have shown that the ratio of carotenoid to chlorophyll content gradually increases during vegetation stress or aging [83].

3.1.1. Descriptive Statistical Analysis of the Measured PPCs in the PeITCs

Figure 4 is the descriptive statistical chart of measured chlorophyll and carotenoid contents in the PeITCs. It can be seen from the figure that the average value of Chl is 2.007 mg/g, the maximum value is 4.813 mg/g and the minimum value is 0.665 mg/g. The mean value of Car is 0.703 mg/g, the maximum is 1.609 mg/g, and the minimum is 0.227 mg/g. The C.V of both are basically the same, and referring to Figure 3, it can be known that they are both of strong variability. The PPC of most Populus euphratica is lower than the average value, and the ratio of chlorophyll to carotenoid is less than 3. In general, in a normal plant, the ratio is about 3:1 [84]. This is a reflection of the low PPC, weak photosynthesis, general health and growth of Populus euphratica in Daliyabuyi Oasis. This outcome is consistent with the statistical results of phenotypic health characteristics of Populus euphratica observed in the field [36,38]. On the other hand, by comparing the Chls and Cars in June and August, it was found that the average contents of chlorophyll and carotenoid in PeITC in June are more than twice that in August, while the mean ratio between them was significantly lower in June than in August. This suggests that the overall growth of Populus euphratica is better in June than in August. The photosynthetic intensity of Populus euphratica in August is significantly weakened, but perhaps the environmental stress suffered by Populus euphratica in June is stronger than that in August.

3.1.2. Spatial Distribution Characteristics of the Measured PPCs in the PeITCs

In this study, a preliminary set of health grading criteria for PeITC was established based on the measured PPCs of PeITCs (see

Table 2. Relationship between canopy health degree and canopy PPC in Populus euphratica individual tree.

	Poor Health	Sub-Health	Health	Quality Health
Car (mg/g)	<0.4	≥0.4 and <0.7	≥0.7 and <1.2	≥1.2
Chl (mg/g)	<1.2	≥1.2 and <2.0	≥2.0 and <3.5	≥3.5

). According to the mean value of PPCs, the health status of PeITC could be classified into four levels: poor health, sub-health, health, and quality health. What needs to be specially stated here is that this health grading criteria is only formulated according to the data obtained in this study, and whether it is universal or not needs further study.

Figure 5 shows the spatial distribution map of PeITC health degree, which is based on GF-6 fused true color imagery in the study area. According to the measured carotenoid content of PeITC and Table 2, the canopy health degree was determined. It was made by using different colors to represent the corresponding canopy health degree, and finally inputting the coordinates of the sample trees. Among them, the tree points in 2019 were determined by averaging the Cars of five sample trees in each quadrat. Then the tree points in 2020 that are in the same position as in 2019 were deleted. Therefore, there are 16 tree points in this map. Although no clear macroscopic pattern can be derived from the map, on the one hand, it is caused by the limitation of sample number, on the other hand, the health status of Populus euphratica individual tree is affected by various factors, which is a very complicated problem in itself. From another perspective, the spatial distribution map can be used to prove the rationality of sample selection in this study. It is distributed evenly and widely in the oasis and covers different health degrees, which lays a foundation for the reliability of the subsequent models.

3.1.3. Temporal and Spatial Distribution of Measured PPCs

Wang et al. continuously measured the chlorophyll content of Populus euphratica leaves in field plots from May to October. It was found that the monthly variation characteristics were first decreased, then increased and then decreased. Moreover, the chlorophyll contents of healthy Populus euphratica leaves were generally higher than that of Populus euphratica under water stress. Besides, the ratio of chlorophyll a to chlorophyll b in healthy Populus euphratica leaves was significantly lower than that under water stress from June to September [51]. In this study, the temporal variation of the measured chlorophyll and carotenoid contents in PeITCs showed that June was greater than August, while the ratio of chlorophyll to carotenoid in June was less than that in August. This result is in general agreement with the findings of Wang et al. On the other hand, the present study did not obtain similar results as Wang et al. in the ratio of chlorophyll a to chlorophyll b. No matter in June or August, the chlorophyll a:b value in this paper increased abnormally, and even chlorophyll b was not detected in some samples. The reason for this phenomenon may be that due to sufficient light in Daliyabuyi Oasis, Populus euphratica does not need chlorophyll b, a pigment used to further capture and transmit light energy, but carotenoids, a pigment used to prevent light damage.

In addition, the present study failed to produce similar results to Wang et al. [85] regarding the spatial distribution characteristics of the measured chlorophyll and carotenoid contents in PeITCs. There was no significant law that the PPCs decreased gradually with the degree of water stress (from the south to the north of the oasis). Even in the northern part of the oasis, where water was supposed to be extremely scarce, high PPCs were observed in PeITCs. The causes of this situation can be very complicated. Individual differences of Populus euphratica are influenced by many factors. Poor growth may exist in a good environment, while good growth may also exist in a bad environment. The specific cause is still unclear and further research is needed. Based on the current fieldwork in Daliyabuyi Oasis, it is speculated that Populus euphratica adapts to the extreme arid environment by reducing the number of leaves and increasing the PPCs in leaves. It could also be due to a subjective bias in sample tree selection, where good growth was selected with a subjective bias in a poor environment.

Finally, it is important to note that: (1) this study did not separately analyze the measured PPC data of PeITC in 2019 and 2020. Instead, the data of carotenoid content for these two years were blended to visualize their spatial distribution characteristics in order to echo the basis of model construction. (2) In this paper, when the PPC of leaves was converted into the PPC of canopy, it was simply averaged for convenience. However, the reality is not so straightforward, and more influencing factors should be considered in the follow-up research. For example, the influence of canopy leaf area index [55] should be considered to make the result of canopy PPC more scientific. (3) The health grade classification based on PPC in this paper is still very immature. It needs further research and discussion to build a more objective and accurate health evaluation system.

3.2. Extraction of the PeITCs

3.2.1. Extraction of Populus Euphratica Crown Regions Based on GWO-OTSU Multi-Threshold Segmentation Algorithm

Because the single-band image is the necessary input condition for threshold segmentation, it is imperative to transform the multispectral image into a single-band vegetation index image that can highlight Populus euphratica. Thanks to the four-band information of multispectral UAV images, a lot of spectral vegetation indices can be calculated. The appropriate vegetation index can improve the accuracy of threshold segmentation. This makes the choice of which vegetation index to use as the base map for threshold segmentation an issue worth considering. The spectral indices that can be calculated for all four bands were calculated in this study (see

Table 3. The broad-band spectral indices used in this study.

Spectral Indices	Calculation Formula	References
Burn Area Index (BAI)	$\frac{1}{{(0.1-Red)}^2+{(0.06-NIR)}^2}$	[86]
Difference Vegetation Index (DVI)	$NIR-Red$	[87]
Green Chorophyll Index (GCI)	$(\frac{NIR}{Green})-1$	[88]
Green Difference Vegetation Index (GDVI)	$NIR-Green$	[89]
Global Environment Monitoring Index (GEMI)	$\begin{array}{l}GEMI=eta(1-0.25\times eta)-\frac{Red-0.125}{1-Red}\\ eta=\frac{2\times (NI{R}^2-Re{d}^2)+1.5\times NIR+0.5\times Red}{NIR+Red+0.5}\end{array}$	[90]
Green Normalized Difference Vegetation Index (GNDVI)	$\frac{(NIR-Green)}{(NIR+Green)}$	[91]
Green Optimized Soil Adjusted Vegetation Index (GOSAVI)	$\frac{NIR-Green}{NIR+Green+0.16}$	[92]
Green Ratio Vegetation Index (GRVI)	$\frac{NIR}{Green}$	[92]
Green Soil Adjusted Vegetation Index (GSAVI)	$1.5\times \frac{(NIR-Green)}{(NIR+Green+0.5)}$	[92]
Infrared Percentage Vegetation Index (IPVI)	$\frac{NIR}{NIR+Red}$	[93]
Modified Chorophyll Absorption Ratio Index-Improved (MCARI2)	$\frac{1.5\times \left[2.5\times (NIR-Re\mathrm{d})-1.3\times (NIR-Green)\right]}{\sqrt{{(2\times NIR+1)}^2-(6\times NIR-5\times \sqrt{Red})}-0.5}$	[94]
Modified Non-Linear Index (MNLI)	$\begin{array}{l}MNLI=\frac{(NI{R}^2-Red)\times (1+L)}{NI{R}^2+Red+L}\\ L=0.5\end{array}$	[95]
Modified Simple Ratio (MSR)	$\frac{(\frac{NIR}{Red})-1}{(\sqrt{\frac{NIR}{Red}})+1}$	[96]
Modified Soil Adjusted Vegetation Index 2 (MSAVI2)	$\frac{2\times NIR+1-\sqrt{{(2\times NIR+1)}^2-8\times (NIR-Red)}}{2}$	[97]
Modified Triangular Vegetation Index (MTVI)	$1.2\times \left[1.2\times (NIR-Green)-2.5\times (Red-Green)\right]$	[94]
Modified Triangular Vegetation Index-Improved (MTVI2)	$\frac{1.5\times \left[1.2\times (NIR-Green)-2.5\times (Red-Green)\right]}{\sqrt{{(2\times NIR+1)}^2-(6\times NIR-5\times \sqrt{Red})-0.5}}$	[94]
Normalized Difference Vegetation Index (NDVI)	$\frac{(NIR-Red)}{(NIR+Red)}$	[92]
Non-Linear Index (NLI)	$\frac{NI{R}^2-Red}{NI{R}^2+Red}$	[98]
Optimized Soil Adjusted Vegetation Index (OSAVI)	$\frac{(NIR-Red)}{(NIR+Red+0.16)}$	[99]
Renormalized Difference Vegetation Index (RDVI)	$\frac{(NIR-Red)}{\sqrt{(NIR+Red)}}$	[100]
Red Green Ratio Index (RGRI)	$\frac{Red}{Green}$	[101]
Soil Adjusted Vegetation Index (SAVI)	$\frac{1.5\times (NIR-Red)}{(NIR+Red+0.5)}$	[102]
Sum Green Index (SGI)	$SGI=Green$	[103]
Simple Ratio (SR)	$\frac{NIR}{Red}$	[104]
Transformed Difference Vegetation Index (TDVI)	$1.5\times \left[\frac{(NIR-Red)}{\sqrt{NI{R}^2+Red+0.5}}\right]$	[105]
Triangular Vegetation Index (TVI)	$\frac{120\times (Red edge-Green)-200\times (Red-Green)}{2}$	[106]
Wide Dynamic Range Vegetation Index (WDRVI)	$\frac{(0.2\times NIR-Red)}{(0.2\times NIR+Red)}$	[107]

). By comparing the effect of these 27 spectral indices in highlighting vegetation in each site, it was considered that the NDVI had the best prominence and universality. Hence, the NDVI map was finally selected as the basis for the calculation of the GWO-OTSU method.

Figure 6 is a schematic diagram of the Populus euphratica crown extraction steps in 2019-A quadrat. The extraction process of Populus euphratica crown in other quadrats is the same, so it is not displayed again. Figure 6a and Figure 6b are the false color synthetic multispectral UAV image and NDVI image of 2019-A quadrat, respectively. Figure 6c–f show the extraction process of Populus euphratica crown, and the specific steps mainly include: (1) setting the threshold number of segmentation to 2 and inputting the NDVI map into the GWO-OTSU algorithm for operation. (2) The output result of the GWO-OTSU algorithm was binarized. (3) The enhanced Forst filter was used to perform 7 × 7 filtering for binarization result, and small noise points were filtered. (4) The zero value of filter graph was nulled to facilitate vectorization operation. (5) The vectorization of the null-valued raster image was performed to obtain a vector map of Populus euphratica crowns.

Next, the vector map of Populus euphratica crown should be optimized. First, the vector areas were counted, and vectors with an area less than 3 m² were removed (based on the results of field research on Populus euphratica in the study area, the range of PeITC area was set to 3–80 m²). Second, the extraction result was checked by visual determination and the non-Populus euphratica vectors were removed to ensure the extraction accuracy. Third, all Populus euphratica quadrat images were processed in turn. The optimal threshold results for each quadrat are shown in

Table 4. Table of the optimal threshold distribution obtained by GWO-OTSU double-threshold segmentation for each quadrat.

Plot	Threshold-1	NDVI-1	Threshold-2	NDVI-2
2019-A	52.032	0.204	102.489	0.402
2020-A	57.253	0.225	116.453	0.457
2020-B	60.125	0.236	126.075	0.494
2019-B	42.312	0.166	81.737	0.321
2019-C	41.251	0.162	77.675	0.305
2019-D	56.498	0.222	101.374	0.398
2019-E	56.179	0.220	104.083	0.408
2019-F	50.213	0.197	99.935	0.392
2019-G	36.491	0.143	75.350	0.295

, and the vector extraction results of Populus euphratica canopy in each quadrat are shown in Figure 7.

The multispectral UAV images have more advantages in spectral information compared with ordinary RGB UAV images, which is convenient to further improve the accuracy of feature classification and extraction. Some scholars have carried out tree canopy extraction studies based on multispectral UAV orthophotography [108,109,110]. The results of their studies all proved that multispectral UAV orthophotos can effectively and accurately extract tree canopies. In this study, the spectral differences between Populus euphratica, Tamarix chinensis, soil and so on were studied by using the multispectral UAV orthophoto images. According to local conditions, a two-threshold segmentation algorithm for fast extraction of Populus euphratica canopies was proposed, which solved the automatic extraction problem of Populus euphratica canopy more simply and efficiently. Of course, under some special circumstances, such as when the vegetation index values of Populus euphratica are very similar to those of Tamarix chinensis and Phragmites communis, the multithreshold segmentation method may not be able to extract Populus euphratica completely and accurately. Besides, generally speaking, multithreshold segmentation method can only be used for extraction in small regions.

3.2.2. Segmentation of the PeITCs Based on Multiresolution Segmentation Algorithm

With the increase of high spatial resolution images, individual tree canopy segmentation has gradually become an important research direction in the field of forestry remote sensing. Currently, there are two main types of individual tree canopy segmentation methods. One is to detect the spectral or height maximum within the window as the canopy center point, and then use the center point as a reference to detect the canopy, such as the local maximum method. The second is canopy detection based on canopy contours, such as the multiresolution segmentation method. Because Populus euphratica and other broad-leaved tree species often have multiple highest points in one crown. The first method is more suitable for pure coniferous forest areas with distinct peaks. The second method is applicable to both coniferous and broad-leaved forests. Moreover, in view of the irregular shape and size of broad-leaved forest crowns, multiresolution segmentation method can extract crowns of different sizes with different scales [111]. Zhou and Zhang obtained images of Larix gmelinii and Pinus tabulaeformis at different forest ages by UAV oblique photography and used multiresolution segmentation method to segment individual trees, which achieved good results [112]. In this paper, the multiresolution segmentation method based on multispectral UAV images could give full play to the advantages of image spectral information and produce better results. However, the shortcoming of multiresolution segmentation method is that it is affected by subjective factors and cannot be fully automated.

The results in Figure 7 were sequentially input into eCognition Developer 10.2 software to perform multiresolution segmentation. After repeated tests, the best segmentation effect of PeITCs was obtained when the segmentation scale was 120, the shape index was 0.7, and the compactness index was 0.5. Then, the vector results of PeITCs were visually optimized for each quadrat. The vectors with an area less than 3 m² were deleted again, and the Populus euphratica canopy vectors with an area greater than 80 m² were further segmented manually to ensure the rationality and accuracy of PeITC extraction. The final results are shown in Figure 8.

Apart from that, because the colored board was not placed next to the isolated Populus euphratica single tree, to further verify the extraction accuracy of the PeITCs, here, the visual interpretation and manual vectorization were applied for verification. In ArcGIS 10.8, 30 sample trees were individually vectorized manually, and the average NDVI values of 30 sample trees were calculated by using the zonal statistics tool. Then, the vectors of these 30 sample trees were extracted separately from the multiresolution segmentation results. The average NDVI values of these vectors were also calculated using the zonal statistics tool. Afterwards, the RMSE and R values of the mean NDVI values obtained by these two methods were calculated. The results proved to be in good agreement, with RMSE of 0.038 and R reaching 0.951. To some extent, it proves the reliability of PeITC semiautomatic extraction results.

At last, the spectral indices batch tool in ENVI 5.3 was used to calculate 27 spectral indices images of all quadrats, and band synthetic images were output. The band synthetic images and the vector results of PeITCs were input into the zonal statistics tool to calculate the mean values of PeITCs 27 spectral indices, and the statistical results were output as Excel files. In the next step of regression model construction, the data in these excel files would be used as model independent variables.

3.3. Construction of Estimation Model for the PPCs in the PeITCs

In general, there are two kinds of quantitative estimation models: statistical models and mechanistic models [113]. The statistical model is based on spectral indices is simple to use and more resistant to environmental disturbances. It is less influenced by different leaf shapes and canopy structures of Populus euphratica, and related studies have also shown that using spectral indices on multispectral UAV images is better than using reflectance directly [114]. Therefore, a variety of spectral indices and statistical models were used in this paper to seek the optimal universal model for quantitative estimation of PPC in PeITC.

3.3.1. Selection of Independent Variables in Regression Model

The statistical regression model was established based on the measured PPC of 30 sample trees in the 2019-A, 2019-B, 2019-C, 2019-D, 2019-E, 2019-F, 2020-A and 2020-B quadrats and the mean values of 27 spectral indices corresponding to the sample trees in the multispectral UAV images. First, the model independent variables (i.e., 27 spectral indices) were further screened in order to reduce the model invalid independent variables and improve the efficiency of the regression model. Here, the R between the measured PPCs and the mean values of 27 spectral indices corresponding to the 30 sample trees was calculated for screening purposes. The spectral index independent variables whose correlation coefficient significance level was lower than 0.02 were eliminated and, to make the screening results clearer, only values with correlation less than −0.408 and more than 0.408 are shown in Figure 9. According to the screening results, only 12 of the 27 spectral indices were significantly correlated with Chl and Car at the significance level of 0.02. In particular, there are three interesting points that can be found in the diagram. One is that the maximum R value of chlorophyll is less than the extreme R value of chlorophyll a, and there is no spectral index with the R value of chlorophyll b exceeding 0.02 significance level. Two, the maximum R of SPAD is greater than the maximum R of chlorophyll and carotenoid. Three, the R of each spectral index with Car is greater than its R with Chl.

3.3.2. Partition of Calibration Set and Validation Set in Regression Model

After the model independent variables are preferred, it is also necessary to divide the calibration set and validation set of the regression model. Typically, the ratio of calibration set to validation set should be 3:1. In this study, this ratio was set to 5:1 in an effort to maximize the amount of data in the calibration set to compensate for the disadvantage of less total data and to enhance the robustness of the model as much as possible. In addition, five suitable data were selected from the quadrats in different positions according to five different color plates as model validation set. Figure 10 is a descriptive statistical violin plot of calibration set and validation set. As can be seen from Figure 10, no matter chlorophyll or carotenoid, characteristics of the calibration set are basically consistent with those of the entire set. Moreover, validation sets are also uniformly distributed, which indicates that the partition results of calibration sets and validation sets are reasonable and can be used in the next step of model calibration and validation. In this regard, it should be particularly noted that the accuracy of the calibration set and the validation set were denoted by the calibration coefficient of determination (R²_C), the calibration root mean square error (RMSE_C), the calibration residual prediction deviation (RPD_C) and the validation coefficient of determination (R²_V), the validation root mean square error (RMSE_V), the validation residual prediction deviation (RPD_V), respectively for differentiation.

3.3.3. Estimation Model Accuracy Evaluation and Validation

Then, seven different regression models (PPC as the dependent variable and 12 spectral indices as the independent variables) were used to construct separate models for estimating chlorophyll and carotenoid contents in PeITCs based on the calibration sets. Figure 11 shows the Taylor diagram for the accuracy evaluation of the seven models constructed based on the Car calibration set. It can be seen from the figure that the optimal model is the CART-DT model, with R²_C of 0.843, RMSE_C of 0.084 and RPD_C of 2.525.

Afterwards, the optimal model obtained in the previous step was validated using the validation set. The results are shown in Figure 12. It can be seen from Figure 12 that its R²_V is 0.670, RMSE_V is 0.251, and RPD_V is 1.741. As can be learned from Figure 3, the performance of the model is general but qualified, and it could be used for the preliminary estimation of carotenoid content in PeITC. In addition, the optimal model for estimating Chl was still CART-DT model. Its R²_C and R²_V were 0.709 and 0.495, RMSE_C and RMSE_V were 0.280 and 0.885, RPD_C and RPD_V were 1.854 and 1.406, respectively. Although the chlorophyll model was inferior to the carotenoid model, it was barely qualified and could still be used for the preliminary estimation of chlorophyll content in PeITC in this study.

3.3.4. Inversion of the PPCs in the PeITCs Based on Optimal Model

Finally, the optimal model was used to estimate the PPC in PeITC, and the estimation results were visualized in ArcGIS 10.8. Here, for the sake of saving layout, only the estimation results of chlorophyll content in PeITC are illustrated, as shown in Figure 13.

3.4. Inversion Results Verification of the PPCs in the PeITCs

3.4.1. Verification of Inversion Results Based on GF-6

To verify the reliability of Parrot Sequoia+ multispectral images and the accuracy of canopy PPC model prediction results, GF-6 multispectral imageries were applied for comparative validation to prove the validity of this study.

Table 5. Comparison of bands between GF-6 imagery and Parrot Sequoia+ multispectral UAV image.

Wavelength Range	Parrot Sequoia+			GF-6
	Band Names	Wavelengths	FWHM	Band Names	Wavelengths	FWHM
420–470				Blue	485	70
500–570	Green	550	40	Green	555	70
620–780	Red	660	40	Red	660	60
680–750	Red edge	735	10
780–1500	NIR	790	40	NIR	830	120

lists the wavelength range of each band in general and the band information of Parrot Sequoia+ image and GF-6 imagery. It can be noticed from Table 5 that there are some differences in band settings between Parrot Sequoia+ image and GF-6 imagery. In whole, GF-6 imagery broadband has a wider wavelength coverage but, especially in the setting of the red band, they are similar.

The Parrot RGB image (Figure 14A), Parrot Sequoia+ four-band multispectral image (Figure 14B), and Parrot Sequoia+ NDVI image (Figure 14(C1)) with GF-6 panchromatic-multispectral fusion image (Figure 14(D1)) of the 2019 validation sample (2019-G) and their corresponding PeITC extraction vector diagram (Figure 14(C2,D2)) are shown in Figure 14. Here, owing to the geometric offset between the GF-6 imagery and the Parrot Sequoia+ image, it was required to resample the Parrot Sequoia+ multispectral image to a resolution greater than 20 cm. Then, it was registered with GF-6 multispectral imagery. Image Registration Workflow module was used in ENVI 5.3 for automatic image registration. First, the drone image was set as the base image and the GF-6 imagery was set as the warp image. The main parameters of image registration, seed point generation and matching were all kept the default settings. Then, when setting warping options, the output pixel size was set to 2 m, and other parameters remained default. Finally, the registered GF-6 imagery was output. After registration, the PeITC vectors with an area attribute less than 4 m² in the Parrot Sequoia+ image segmentation vector file were deleted (the single pixel area of GF-6 fusion imagery was 4 m²). Later, vector file was imported as the region of interest with unique ID based on GF-6 imagery. Lastly, this region of interest was exported as the vector result of PeITC extraction in GF-6 imagery. It is not difficult to get the conclusion from Figure 14 that Parrot Sequoia+ multispectral image is not significantly different from GF-6 multispectral imagery, except for a great difference in spatial resolution. Moreover, the vector extraction results of PeITC are basically the same.

Subsequently, in an attempt to further compare Parrot image with GF-6 imagery, the average band reflectance values of 172 PeITCs were separately extracted to make the spectral reflectance graphs (Figure 15). From the shape of band reflectance curve, the band trends of both Parrot and GF-6 images are roughly the same. On the one hand, this shows that there is no problem with the result of threshold segmentation, and the extracted features are indeed at least vegetation. On the other hand, it indirectly proves the reliability of Parrot Sequoia+ multispectral images. In which, by comparing the red band reflectance alone (Figure 15a,b), it can be observed that the average red band reflectance of GF-6 imagery is slightly higher than that of Parrot image. This may be caused by the presence of mixed pixels in GF-6 imagery. Another interesting phenomenon can be detected: as the line colors identified by ID are arranged from small to large area (the area is between 4–80 m²), it can be concluded that the larger the canopy area of Populus euphratica, the smaller the average band reflectance of PeITC. Furthermore, by comparing Figure 15c with Figure 15d, it can be clearly seen that the mean band reflectance curves of the predicted carotenoid content values in PeITCs are basically consistent except the predicted value of 0.901. Especially in the red band, the band reflectance decreases with the increase of carotenoid content in PeITC. This just matches the red valley characteristic of the vegetation, which seems to be able to verify the prediction results of the model to some extent.

In the end, the R between NDVI of Parrot (NDVI_Parrot) and NDVI of GF-6 (NDVI_GF-6), predicted value of Chl (Chl_pre), and predicted value of Car (Car_pre) were analyzed separately for 172 PeITCs. The results are shown in

Table 6. Correlation statistics of 172 PeITCs.

	NDVIParrot
Carpre	0.486 **
Chlpre	0.483 **
NDVIGF-6	0.331 **

** represents correlation at 0.001 significance level.

. The R between Car_pre and NDVI_Parrot is 0.486 at 0.001 significance level, and NDVI_Parrot in turn correlated with NDVI_GF-6 at a value of 0.331. This finding may also prove the reliability of the model prediction result in a certain sense.

As is well known, radiometric correction of images is a very critical step in quantitative remote sensing applications, which is the basis for subsequent quantitative models to be established. There are two main radiometric correction methods: the empirical line method and the solar irradiance measurement method [48]. At present, several studies have been conducted on radiometric correction of Parrot Sequoia+ multispectral UAV images. Tu et al. [115] evaluated the effects of various radiometric correction methods in horticultural environments based on Parrot Sequoia+ multispectral UAV images. It was considered that there were many factors affecting the accuracy of radiometric correction, and the potential solutions included: (1) appropriate panel deployment, (2) site-specific sensor calibration, (3) appropriate bidirectional reflectance distribution function (BRDF) correction. Cubero-Castan et al. argued that the use of radiometric calibration plates for radiometric correction of multispectral UAV images over large areas was impractical and error-prone. Therefore, research on the radiometric accuracy assessment in a targetless workflow using Pix4D software was carried out based on the multispectral images acquired by Parrot Sequoia+ camera. The results showed that the processed reflectance of the Parrot Sequoia+ camera was highly consistent with the real reflectance of the verification site in all bands, and the Rs were better than 0.98. It was further concluded that the UAV system equipped with multispectral camera and illumination sensor was capable of acquiring regional high-precision reflectance images combined with Pix4D software, without the need for a radiometric calibration plate [47]. Franzini et al. evaluated the geometric and radiometric consistency of Parrot Sequoia camera in acquiring multispectral images and discussed its feasibility in precision agriculture. It was found that the geometric consistency of the Parrot Sequoia camera was good (RMSE could be controlled within 0.1 m), while there were great problems with radiometric consistency. There were errors of more than 20% in some bands and spectral indices values, and the comparisons with Sentinel-2 imageries showed generally higher NDVI values in Parrot Sequoia multispectral images. Despite these problems mentioned above, the authors believed that Parrot Sequoia camera had greater potential for agronomic applications, especially since the emergence of new Parrot Sequoia+ camera would bring preferable radiometric correction effect [116]. In other studies, researchers had found that the accuracy obtained by using spectral indices in quantitative remote sensing analysis based on multispectral images collected by Parrot Sequoia+ camera was higher than that by using reflectance directly [48,114].

These studies have demonstrated the reliability of radiometric correction results for the Parrot Sequoia+ UAV system to a great extent. Thus, in this study, the Parrot Sequoia+ multispectral UAV system without a radiometric calibration board was used in conjunction with Pix4D software to obtain the absolute reflectance images of the quadrats. However, there may be some uncertainties due to its initial application in an arid desert riparian forest environment. In this study, there were still concerns about its adaptability, so the reliability of UAV multispectral images was further verified by using GF-6 multispectral imageries to solidify the research findings. The qualitative validation results proved that the reflectance of UAV multispectral images bands and the estimation results of the model constructed from them were reasonable. In addition, it should be noted that the main reason why only qualitative methods can be used for verification here is that the satellite imageries are all processed by the fusion of panchromatic and multispectral band imageries. This is mainly to improve the spatial resolution of satellite imageries, so that it can be applied to the extraction of PeITC, so as to achieve the purpose of comparison on the single tree canopy scale. However, because the fusion algorithm will more or less change the band absolute reflectance of the original multispectral satellite imageries, it is no longer suitable for quantitative comparison. It is undeniable that the reflectance results of multispectral drones can still be affected by many factors. After all, the qualitative verification method has its limitations, and it is still the best scheme to compare with the quantitative method. It is better to find a way to carry out quantitative verification in the follow-up. In any case, it is believed that the efforts made in this study to evaluate the effects of multispectral UAV images radiometric correction can further dispel other researchers’ concerns regarding the application of consumer multispectral UAV system in the field of quantitative remote sensing.

3.4.2. Verification of Inversion Results Based on ZY1-02D

Figure 16 shows the schematic diagram of 2020-C verification quadrat. Five typical Populus euphratica were evenly selected from this quadrat, and their peripheries were marked with five colored plates of white, black, yellow, red, and blue to ensure that the five verified Populus euphratica could be distinguished in the UAV image. The Parrot Sequoia+ UAV system was then manipulated to gain multispectral UAV images of this verification quadrat. Next, the SPAD mean values of those five PeITCs were measured and recorded in sequence. In the last part, the field photos of the five verified Populus euphratica were taken and the canopy relative health degrees were visually judged. By manual vectorization, the CART-DT model was utilized to calculate the canopy PPC of those five verified Populus euphratica, and the model prediction results were obtained.

To begin with, the R values between the 46 measured PPCs and the measured SPAD were calculated. They were used to determine whether the SPAD data could effectively validate the model prediction results. According to the statistical results in Figure 17a, it can be seen that SPAD is significantly correlated with Chl and Car at the significance level of 0.001 (0.001 significance test value of 46 data: 0.460). The R between SPAD and Chl is 0.684, and the linear fitting R² is 0.468 (Figure 17b). This result showed that SPAD could not well characterize the Chl. There was still some uncertainty in using SPAD to validate the model prediction results. Consequently, this study attempted to use ZY1-02D fused hyperspectral imagery as another validation basis to further verify the reliability of the model prediction results.

Because the spatial resolution of ZY1-02D fusion imagery was 2.5 m, the segmentation effect of PeITCs was not satisfactory. Thereupon, a pixel as pure as possible was selected in the hyperspectral imagery to represent each of these five verified PeITCs. It can be recognized from the pixel spectral curve in Figure 18a that this is not a standard vegetation spectral curve, but a mixed pixel spectral curve. However, vegetation accounts for the majority, so it can still be used as the basis for judgment. In particular, it is known from the results of previous studies that, in general, there is a significant correlation between red edges and vegetation Chl [117]. Here, to qualitatively compare the model prediction results with the ZY1-02D hyperspectral imagery spectral results, the red edge may be a plausible basis for judgment. The range of red edge is 680–750 nm [118], and the description of red edge includes the position and slope of red edge. The greater the Chl, the greater the slope of the red edge, and the slope is the first order derivative of the spectral curve. Figure 18c is the first derivative curve of the five PeITCs. For a clearer comparison of the spectral curve and slope curve of the five verified PeITCs in the red edge range, in Figure 18, this section of the curve is intercepted respectively, and the results are shown in Figure 18b,d.

Meanwhile, the Carotenoid Reflection Index 2 (CRI2) values of five verified PeITCs were calculated by ZY1-02D hyperspectral imageries to further increase the verification basis of the model prediction results. Higher CRI2 values mean greater carotenoid concentration relative to chlorophyll [119].

Table 7. Validation table of estimation accuracy.

PeITC	SPAD	Carpre	Chlpre	ZY1-02Dred edge	ZY1-02DCRI2
White board	33.8	0.396	1.062	0.0031	0.846
Yellow board	34.3	0.411	1.223	0.0024	0.765
Blue board	40.5	0.562	1.610	0.0025	0.778
Black board	47.3	0.562	1.610	0.0027	0.740
Red board	47.5	0.411	1.223	0.0025	0.829

lists mean SPAD, Chl_pre, Car_pre, red edge value calculated by ZY1-02D (ZY1-02D_{red edge}) and CRI2 value calculated by ZY1-02D (ZY1-02D_CRI2) values respectively.

At last, to qualitatively compare the predicted results of the model, measured SPAD, and ZY1-02D hyperspectral imagery, the concept of relative health degree was introduced here, and the canopy health status of Populus euphratica could be divided into three grades: good (3), fair (2) and poor (1). According to the numerical values of different indicators, the relative health of five verified PeITCs could be classified into three grades. Of them, the visual judgment of relative health degree of PeITC was mainly based on the green to withered ratio and greenness. The comparative results of relative health are shown in Figure 19. Each verified PeITC has the same relative health of at least three indicators, and almost all of them have one indicator with the same relative health of Chl. To a certain extent, this proves the scientificity of the model prediction results.

Compared with the findings of Kopačková-Strnadová et al. [23], in this paper, there are many similarities with them. The results of this study are reliable and reasonable, and thanks to the systematicity of the model construction and verification process, the CART-DT model used in this paper is superior in quantitative estimation accuracy, which may be related to the matching between the prediction mechanism of CART-DT model and the problems studied in this paper. It can basically meet the precision requirement of quantitative estimation of PPC in PeITC. The feasibility of multispectral UAVs in quantitative remote sensing for forestry can be further affirmed but, no matter how, it is still disturbed and limited by many elements. The accuracy of the model is still not very high, and GF-6 and ZY1-02D imageries can only be used for qualitative verification, so these need to be gradually improved in the following research.

4. Conclusions

In this paper, an exploratory study was carried out on whether the PPCs in the PeITCs could be quantitatively retrieved from the multispectral UAV images, a quantitative remote sensing issue of forestry in arid area. The main findings include:

(1)

The average value of Chl in Daliyabuyi Oasis was 2.007 mg/g, the maximum value was 4.813 mg/g and the minimum value was 0.665 mg/g. The mean value of Car was 0.703 mg/g, the maximum was 1.609 mg/g, and the minimum was 0.227 mg/g. The C.V of both were basically the same and they were both of strong variability. PPCs were generally low, reflecting the general health level of Populus euphratica. The average contents of chlorophyll and carotenoid in PeITC in June were more than twice that in August, while the mean ratio between them was significantly lower in June than in August. This suggests that the overall growth of Populus euphratica was better in June than in August. The photosynthetic intensity of Populus euphratica in August is significantly weakened, but perhaps the environmental stress suffered by Populus euphratica in June was stronger than that in August. The measured PPCs had no obvious spatial distribution law. However, from another perspective, it was distributed evenly and widely in the oasis and covered different health degrees, which lays a foundation for the reliability of the subsequent models.

(2)

By comparing the effect of 27 spectral indices in highlighting vegetation in each site, it is considered that the NDVI had the best prominence and universality. Based on GWO-OTSU multithreshold segmentation method, the canopy area of Populus euphratica could be quickly and effectively extracted from the quadrat NDVI map. The multiresolution segmentation method could also accurately segment the PeITCs from the Populus euphratica crown region, and the best segmentation effect of PeITCs was obtained when the segmentation scale was 120, the shape index was 0.7, and the compactness index was 0.5. The visual interpretation and manual vectorization of 30 sample trees were applied for verification. The RMSE and R values of the mean NDVI values obtained by these two methods were calculated. The results proved to be in good agreement, with RMSE of 0.038 and R reaching 0.951. To some extent, this proves the reliability of PeITC semiautomatic extraction results.

(3)

According to the model independent variables screening results, only 12 of the 27 spectral indices were significantly correlated with Chl and Car at the significance level of 0.02. No matter chlorophyll or carotenoid, characteristics of the calibration set were basically consistent with those of the entire set. Moreover, validation sets were also uniformly distributed, which indicates that the partition results of calibration sets and validation sets were reasonable and could be used in the next step of model calibration and validation. The CART-DT model performed best in the estimation of the PPCs in the PeITCs, in which, when estimating the Car, the R²_C was 0.843, the RMSE_C was 0.084, the RPD_C was 2.525, the R²_V was 0.670, the RMSE_V was 0.251, the RPD_V was 1.741.

(4)

Qualitative comparison of spectral reflectance and NDVI values between GF-6 multispectral imagery and Parrot Sequoia+ multispectral images on the 172 PeITCs showed the reliability of Parrot Sequoia+ multispectral images and the accuracy of the prediction results of the canopy PPC model to some extent. The comparison results of five PeITCs’ relative health degree judged by field vision judgment, measured SPAD value, Chl_pre, ZY1-02D_{red edge} and ZY1-02D_CRI2 in 2020-C quadrat further prove the scientificity of inversion model.

Thus, it was proved that Parrot Sequoia+ multispectral UAV images can be applied for quantitative inversion of PPC in PeITC and health evaluation of Populus euphratica, which made up for the lack of research in this field and further affirmed the serviceability of multispectral UAV in forestry quantitative remote sensing application.

In the next study, it is imperative to further increase the sampling volume and optimize the sampling scheme, and to further improve the accuracy and automation of the PeITC extraction by means of UAV oblique photography or LiDAR technology where possible. What is more, perhaps a physical model could be tried to construct a PPC estimation model. Finally, health assessment of individual Populus euphratica trees can be carried out in combination with other health indicators.