Research Progress on the Early Monitoring of Pine Wilt Disease Using Hyperspectral Techniques.

Pine wilt disease (PWD) caused by pine wood nematode (PWN, Bursaphelenchus xylophilus) originated in North America and has since spread to Asia and Europe. PWN is currently a quarantine object in 52 countries. In recent years, pine wilt disease has caused considerable economic losses to the pine forest production industry in China, as it is difficult to control. Thus, one of the key strategies for controlling pine wilt disease is to identify epidemic points as early as possible. The use of hyperspectral cameras mounted on drones is expected to enable PWD monitoring over large areas of forest, and hyperspectral images can reflect different stages of PWD. The trend of applying hyperspectral techniques to the monitoring of pine wilt disease is analyzed, and the corresponding strategies to address the existing technical problems are proposed, such as data collection of early warning stages, needs of using unmanned aerial vehicles (UAVs), and establishment of models after preprocessing.

1. Introduction

Pine n class="Disease">wilt disease (pan> class="Gene">PWD) is a major quarantine disease that causes the devastating death of Pinus species due to the pine wood nematode (PWN) Bursaphelenchus xylophilus (Steiner and Bührer) Nickle [1,2,3,4]. Because of its rapid onset and spread in addition to the resulting high mortality, this disease is very difficult to control [5]. Currently, the disease is widely distributed in eight countries, namely, in North America: the United States, Mexico, and Canada; in East Asia: China, Japan, South Korea, and North Korea; and in Europe: Portugal. However, globally, 52 countries have declared B. xylophilus as a quarantine pest and PWD as a quarantine disease [2,3,6,7,8,9,10,11,12]. Among the affected countries, China, Japan, and South Korea are the most severely impacted. In 1979, large numbers of pine trees died in Japan, with numbers of dead trees reaching as high as 2.43 million m3 [13]. In South Korea, approximately 78.11 km2 of pine forest was destroyed in 2005 [14]. Since PWD was first described in the Sun Zhongshan Mausoleum in Nanjing, China, in 1982, the disease has rapidly spread in the tropical and subtropical regions of China in only a few decades. In recent years, PWD has gradually invaded warm temperate zones and exceeded the 10 °C average annual temperature [15].

According to the n class="Species">Chinese Forestry Apan> class="Disease">dministration’s No. 4 National Pine Wood Nematode Epidemic Area Announcement in 2019, PWD is widely distributed in 18 provinces [16]. Many trees have been destroyed, leading to large-scale deaths of pine trees, which has severely disrupted the ecological forestry environment and resulted in devastating losses. Because pine trees are environmentally friendly and valuable to the economy, the Chinese government, in addition to numerous others, attaches great importance to this resource, and have strengthened efforts toward the prevention and control of PWD, especially in recent years. Our group has done a lot of work in this field. A wood sampling method was reported in 2010 [17]. It can estimate nematode numbers rapidly in the wood of Pinus massonina with only 10 min. In 2011, our group developed a rapid detection of B. xylophilus in stored Monochamus alternatus. This method can detect nematodes by PCR amplification [18]. One of the key strategies in combatting PWD is the early identification of epidemic areas via large-scale forest monitoring. PWD symptoms can be leveraged for the intelligent monitoring of forest pines [4]. Large-scale forest monitoring of pine wilt disease can be achieved by hyperspectral techniques.

n class="Disease">Hyperspectral techniques canpan> be used for object identificationpan> anpan>d object character anpan>alysis. It is widely used in agriculture, oceanpan>ography, geology, atmosphere, anpan>d environpan>mental sensing [19]. In agriculture, hyperspectral techniques can be used for pest and disease monitoring [20,21,22,23,24,25], biomass estimation [26], and biological stress monitoring [27]. In the field of pest and disease monitoring, it has been applied to many plants, including tomatoes, grapes, and wheat [20,21,22]. Similarly, it can also be applied to the early monitoring of pine wilt disease. However, because hyperspectral techniques are not mature, there are a lot of problems and difficulties that limit its practical applications. Therefore, in order to inspire other research in this area, this review discusses the progress, challenges, and suggestions on the early monitoring of pine wilt disease using hyperspectral techniques.

2. PWD Symptoms

PWD is generally tranpan>smitted by insect vectors such as pan> class="Species">M. alternatus. When insect vectors feed on pine trees, the pine wood nematode parasite is transferred to pine trees, and PWD caused by the nematode is generally reflected in the canopy of the host plant. PWD is generally divided into three stages on the basis of its symptoms [28]:

(1) The difference between a healthy plant and plant during an early disease stage cannot be observed with the naked eye. The duration of this early stage is mainly affected by pine age and species, environmental temperature, and n class="Species">pine wood nematode virulence. pan> class="Species">Pine wood nematodes begin to multiply in the xylem of pine trees, blocking the internal ducts of pine trees. Pine trees cannot perform normal water metabolism, and pine resin secretion is reduced along with slight changes in chlorophyll and carotene contents. Clevers’ research showed that the red-edge chlorophyll index is linearly correlated with the canopy chlorophyll content [29]. In addition, Köksal found that the normalized difference vegetation index is strongly associated with the carotene content of leaves [30]. Therefore, the change in chlorophyll and carotene content can be detected by hyperspectral remote sensing.

(2) In the middle stage of the disease, pine tree resin secretion completely stops, anpan>d the tranpan>spiration effect is weakened. Needles in the crown of the tree gradually become yellowish. Visual differences between healthy anpan>d pan> class="Disease">infected pines are evident at this stage. The wilting appearance of the pine during this period differs from that of the healthy pine. PWD causes pine needles to wilt from the inside to the outside.

(3) At the end stage of the disease, needles on the crown of the tree turn brown or reddish brown. The whole plant dies, and the needles do not fall that year.

The spectral characteristics of pines vary according to the type of stress. Different spectral indices canpan> be used to build models to identify pan> class="Gene">PWD and differentiate it from other decline agents, such as bark beetles, defoliating fungi, or even severe droughts [31].

3. Traditional PWD Monitoring Technology

3.1. PWD Pine Tree Monitoring Methods

n class="Gene">PWD diagnpan>osis depends onpan> pan> class="Species">human reconnaissance and visual rating (Figure 1), followed by the isolation of nematodes found in diseased wood. Microscopy is used to observe the morphological characteristics of the nematodes and determine whether PWNs are present in the pine tree. Traditional methods of morphological identification require training for recognizing nematodes. Molecular identification is also available but is expensive for large-scale monitoring [32].

Figure 1

Pine seedlings: healthy plus five different infection levels.

In certain harsh conditions that make it difficult to enter a forest, the transmission source cannot be treated in time after pine trees are n class="Disease">infected with nematodes, resulting in large-scale damage to the pine forest. In additionpan>, the field diagnpan>osis of pan> class="Gene">PWD based on the analysis of visual symptom is subject to variation due to the different evaluators, and it is easy to draw incorrect conclusions [33,34,35,36]. Many researchers have developed DNA-based methods to rapidly identify PWD using molecular biology methods [37,38,39]. For example, a sampling-staining-PCR detection workflow system was established for the first time by Wang and colleagues [40]. This system can detect pine wood nematodes directly from infected P. massoniana within 5 h without requiring nematodes to be isolated from samples. However, this method does not address the problem of allowing pine wood nematode diagnosis before disease symptoms appear (Figure 1).

3.2. Pine Wood PWD Monitoring Technology

n class="Gene">PWD canpan> be monpan>itored with the “punching anpan>d flowing resin” method since this disease leads to a reductionpan> in resin secretionpan> from pine trees. Because this method is not specific to pan> class="Gene">PWD, it was gradually eliminated [41]. On the other hand, PWD can also be diagnosed by monitoring the density of M. alternatus. M. alternatus is abundant in pine forests, but this insect does not always transmit the PWN responsible for PWD. In fact, its odds of carrying PWN range from 0 to 60%, and few of these PWNs successfully invade tissue. Therefore, monitoring PWD on the basis of M. alternatus is ineffective [42].

Traditional methods for monitoring n class="Gene">PWD are time-consuming anpan>d labour-intensive. This disease canpan> be diagnosed with certainty by nematode separation anpan>d morphology identification. Therefore, anpan> efficient method is urgently needed for the intelligent monitoring of pan> class="Gene">PWD.

4. Principles of Monitoring PWD with Hyperspectral Remote Sensing Technology

The advantage of using hyperspectral remote sensing technology in disease classificationpan> is that it canpan> combine image informationpan> anpan>d spectral informationpan>. Image informationpan> canpan> be used to anpan>alyse surface quality features such as sample size, profiles, anpan>d defects. Because spectral absorptionpan> varies from onpan>e componpan>ent to anpan>other, anpan> image reflects a certain feature at characterised banpan>ds, anpan>d spectral informationpan> canpan> fully reflect a discrepanpan>cy in the physical structure anpan>d chemical compositionpan> of a sample [43]. These characteristics determine the unique advanpan>tages of spectral imaging technology in testing the internal anpan>d external quality of farm products. Spectral imaging technology makes full use of the absorptionpan> or radiationpan> characteristics of a substanpan>ce in different electromagnpan>etic spectra. One-dimensionpan>al spectral informationpan> is added onpan> the basis of ordinary two-dimensionpan>al space imaging. Due to differences in the compositionpan> of various substanpan>ces, differences also exist between their corresponpan>ding spectra such that the spectrum of the target canpan> be used in identificationpan> anpan>d classificationpan>. To achieve monpan>itoring, planpan>t diseases are identified through various parameters, such as morphological chanpan>ges, temperature chanpan>ges, tranpan>spirationpan> rates, anpan>d volatile organpan>ic compound release by pathogen-pan> class="Disease">infected plants [44]. The delicate process of host–pathogen interactions also affects the optical characteristics of plants to a certain extent [45]. Data about these changes can be obtained by sensors. Hyperspectral sensors offer the most abundant data information. This technology can supply sufficient evidence, demonstrating its feasibility for the early detection of plant diseases. With the development of hyperspectral technology in recent years, increasing numbers of researchers have applied it to plants, including for plant-nutrition monitoring [46], the automatic classification and monitoring of plant diseases [47,48], and their early detection [45], as well as to obtain digital phenotypes for disease-resistant plant breeding [49,50]. For example, pine wilt disease was diagnosed using hyperspectral methods by Kim’s group. In their research, significant differences were observed in vegetation indices such as the normalized difference vegetation Index (NDVI), the green normalized difference vegetation index (GNDVI), the plant senescence reflectance index (PSRI), and the pigment specific normalized difference (PSND), beginning on 20 August 2012 [51]. Our group study the hyperspectral characteristic of healthy and withered masson pine, and Figure 2 shows the research results. If a plant is infected with PWD, the green needles of the pine trees begin to change colour. The spectral reflectance of pines varies over the different degrees of infection. Therefore, disease level can be evaluated on the basis of spectral reflectance.

Figure 2

Detection of pine wilt disease using hyperspectral techniques (a) Hyperspectral images of diseased pine needles (b) Hyperspectral reflectance of healthy and withered masson pine.

Spectral imaging technology can collect many narrow and continuous n class="Disease">hyperspectral reflectanpan>ce datasets in the ultraviolet, visible, near-infrared, anpan>d mid-infrared regions of the electromagnetic banpan>d, providing complete anpan>d continuous spectral curves for each pixel. Imaging spectrometers canpan> be divided according to their resolution, as follows [52].

Multispectral imagers (MSIs): the obtained target has a wavelength band between 320 and 1300 nm, and the spectral resolution is generally close to 100 nm; this technology is mainly used in zone classification [53].

n class="Disease">Hyperspectral imagers (HSIs): the obtained target banpan>d is between 350 anpan>d 2500 nm, anpan>d the spectral resolution is approximately 10 nm; this technology is widely used in various fields, such as food safety, quality, anpan>d biomedical agricultural product testing [54].

Ultran class="Disease">hyperspectral imagers (USIs): the obtained target banpan>d is between 1000 anpan>d 10,000 nm, anpan>d the spectral resolution is below 1 nm; this technology is commonly used when fine detection is required, such as in the case of atmospheric particle detection [55].

5. Hyperspectral Technology in Monitoring of PWD in Forests

5.1. Forest Surveillance of Dead Pine Wilt Disease

As early as 1994, Wu and colleagues, who monitored the n class="Species">Chinese pine caterpillar anpan>d pan> class="Species">pine wood nematode (Bursaphelenchus xylophilus), began to combine high-quality super VHS (video home system) camera systems with global positioning system (GPS) navigation, image processing, and geographic information systems. The captured image data could be converted into digital data for further analysis [56]. Methods based on the Korea Mahalanobis distance (MD) and maximum likelihood were used to classify TM (thematic mapper) and IKONOS images in 1998. In addition, the GIS (geographic information system) spatial analysis function was combined with ground survey data information to predict disasters [57].

The Mahalanobis distance (MD) is defined as follows [58]: where D is the Mahalanobis distance between test samples and reference samples, µ is the mean vector of the reference samples, and S is the covariance estimate of the reference samples.

Maximum likelihood is determined as follows [59]:

Unmanned aerial vehicles (n class="Chemical">UAVs) canpan> carry sensors to collect data in inaccessible mountains anpan>d obtain higher-resolutionpan> image data via software stitching. Figure 3 shows pine forest photographed by unmanpan>ned aerial vehicle. Huanpan>g anpan>d colleagues used small, fixed-wing dronpan>es to monpan>itor anpan> epidemic area with a total area of 145.29 km2 anpan>d automatically identified 1486 dead pine trees. The distributionpan> maps anpan>d coordinate-point locationpan>s of these dead trees were also obtained. The monpan>itoring accuracy exceeded 80%, anpan>d the point accuracy reached 2–3 m [60]. Low-cost, small pan> class="Chemical">UAV high-resolution images were used by Li and colleagues to locate a single dead tree with an accuracy of 83%, which offered technical support for the quick identification of suspected dead trees and the determination of the outbreak area [61]. The hue saturation value (HSV) threshold method was used by Tao and colleagues to extract discoloured pine trees with an accuracy of 60–65%, which improved the efficiency of manual interpretation. This method is suitable for the late monitoring of dead pine trees, and lends theoretical and methodological support for monitoring discoloured pine trees on the basis of drone images [62]. The above studies are based on the RGB three-channel sensor in canopy size. However, the visual portion of the electromagnetic spectrum contains only the three bands of red, green, and blue [63]. Due to its high throughput, this sensor is widely used in the field of remote sensing detection, but it also reduces the analysis accuracy. Therefore, it is important to increase the resolution for the early monitoring of PWD.

Figure 3

Pine forest photographed by unmanned aerial vehicle (UAV). PWD: pine wilt disease. (a) Hyperspectral image of PWD; (b) Digital image of PWD.

5.2. Forest Monitoring for Early PWD Detection

The early monitoring of n class="Gene">PWD has always been a topic of conpan>cern. In 2017, a novel method that combined airborne laser scanpan>ning with high-resolutionpan> spaceborne images was used to detect pine trees affected by pan> class="Gene">PWD, and a support vector machine was applied by Takenaka’s team to classify different degrees of pine disease [64]. Du and others used a portable ASD (analytica spectra devices) FieldSpec full range field spectrometer with a band value between 350 and 2500 nm to collect reflection spectra data of pine needles in the field [65]. Kim’s group used a GER 3700 spectrometer to obtain reflections from infected and uninfected trees [31]. A FieldSpec4 (400–1100 nm) field spectrum analyser was used by Zhang’s group to identify needles with different levels of damage in 70 sampling areas [66]. Wang et al. hyperspectrally recorded black pine older than 20 years-initially healthy-from the date of pine nematode inoculation until their death [67]. Ma and colleagues used the ASD FieldSpec3 spectrometer to measure the hyperspectral reflectance of pine trees in three different areas [68]. Huang et al. used a back-mounted ASD FieldSpec Pro FR field spectroradiometer to measure the needle leaf spectrum in a darkroom [69]. The above studies required that needles be removed from the field to a stable indoor environment to measure their spectral reflectance or, alternatively, the selection of a relatively stable environment to conduct this measurement in the field. The sensor used in the device had a high resolution (350–2500 nm, containing hundreds of bands) and obtained information in the three dimensions of X, Y, and spectrum. However, this needle study was limited to the laboratory, a relatively stable condition, providing data that can support the early detection of PWD in the field. Yet many pine forests are located in places that are difficult to reach, and the early detection of PWD must be performed in the field, which poses many challenges for data collection.

6. Drones Equipped with Hyperspectral Sensors to Monitor Forest PWD

Many scientific methods that obtain good results under laboratory conditions often perform poorly in practical applications. In the study of n class="Gene">PWD, when it comes to conpan>ducting field measurements, researchers canpan> onpan>ly measure forests that they canpan> reach anpan>d test under relatively stable measurement conpan>ditionpan>s. This situationpan> presents a notable challenge that limits model applicationpan> in the field. However, solutionpan>s have emerged with scientific advanpan>ces. The combinationpan> of dronpan>es anpan>d n class="Disease">hyperspectral technology in precision agriculture has solved many problems in agricultural production. With the miniaturisation of GPS receivers, inertial navigation systems, computers, and remote sensing sensors, increasing numbers of researchers have begun to use UAVs to build data collection platforms [70]. Drones have many advantages that make them suitable for use in forests. For example, drones offer the advantages of low consumables, low operating costs, and high-intensity data collection [71]. Drones can carry several different types of optical sensors that can be assembled on the basis of the investigated parameters and operated in a certain amount of time. This method allows users to optimise flight time to avoid inconsistent light intensity caused by clouds [72], and data can be accurately collected when a pest or wildfire outbreak occurs [71,73]. Finally, drones can collect data at high spatial resolution, which is a key aspect of the early monitoring of PWD. Hence, the advantages of drones are evident. However, because of their small size, the flight endurance time of drones is low, which limits their ability to obtain data over a large area [74]. Payload capacity limits the ability to carry airborne sensors, especially technologies that are not sufficiently miniaturised. In addition, the size and shape of the fuselage influences the installation of sensing equipment [70,71,72]. Additionally, it might take a long time to set up a UAV system, which includes the selection process for aircraft and sensors as well as processes for their installation and adaptation. Finally, it is necessary to conduct test flights in the measured environment.

In the selection of drones, eBee-type microdrones are a type of light fixed-wing drone that can operate completely autonomously, and require almost no driving skills [75]. An eight-rotor remote control drone (Cinestar-8 MK Heavy Lift, Freefly Systems, Ren class="Disease">dmonpan>d, WA, USA) was used in more accurate fixed-point acquisitionpan> by Severtsonpan> anpan>d others [76]. The multirotor designpan> of this dronpan>e supplies stable acquisitionpan> space. Dash’s group used coaxial four-wing dronpan>es to collect multispectral images for the monpan>itoring of leaf discolourationpan> caused by simulated forest disease outbreaks [77]. Sanpan>dino used a six-rotor DJI S800 pan> class="Chemical">EVO UAV equipped with a Headwall Nano-Hyperspec® hyperspectral camera to collect data that were used to study the harm caused by myrtle rust to paper tea grown near New South Wales, Australia [78].

7. Prospects of PWD Monitoring Using Hyperspectral Technology

7.1. Hyperspectral Technology in PWD Forest Monitoring

Three different types of sensors are used to study hyperspectral features: RGB, multispectral (MS), anpan>d hyperspectral (HS) sensors [79]. A hyperspectral sensor contains hundreds of bands in the electromagnetic region of 350–2500 nm, with only a few nanometres in each band; thus, consistent and accurate changes in spectral characteristics can be displayed on a spectrogram. Because of the objectivity, continuity, and accuracy of hyperspectral imaging technology, it can be used in automated systems [34]. Therefore, its combination with drone technology allows the possibility of monitoring areas that are out of human reach.

7.2. Hyperspectral Data Acquisition

n class="Disease">Hyperspectral sensors canpan> be used to acquire hyperspectral data, and they are divided into two main types: nonimaging [80,81] and imaging sensors. Nonimaging hyperspectral sensors cannot obtain information on spatial dimensions, and can only measure the average spectral reflectance of the region of interest. Imaging sensors are further divided into three types: push, broom, and filter-based sensors [64,82].

n class="Disease">Hyperspectral imaging sensors must capture onpan>e-dimensionpan>al spectral target informationpan> while capturing two-dimensionpan>al spatial image informationpan>. However, the array of illuminationpan> sensors canpan> onpan>ly detect informationpan> in two dimensionpan>s. Stirring broom anpan>d push-type scanpan>ners eliminate this difference by simultanpan>eously capturing spectral informationpan> of lines or points onpan> the measured object [81]. The object of interest is scanpan>ned by moving or rotating it to form a hyperspectral image. The advantage of these two hyperspectral sensors is that they can supply fine spectral resolution and collect more effective information. On the other hand, these sensors also have certain disadvantages. Due to the speed limitation of target movement or rotation, and the different size of the collection area, it takes a long time to collect images, which limits applications of this approach in agriculture. Filter-based hyperspectral sensors require filter changes during the image acquisition process. The filter allows only specific bands to pass [83]. The image acquisition speed of this kind of sensor is obviously faster than that of the other two types of imaging sensors, but images obtained at different times are inconsistent. This type of camera is known as a fast camera, and if it can simultaneously captures all the bands, the entire image cube is recorded without the scanning process within the same exposure time [84]. Recently, hyperspectral/multispectral snapshot sensors have been introduced in fast cameras. These units use the same mosaic principle as ordinary RGB cameras, and offer faster image acquisition speeds than those of the original hyperspectral imaging sensors.

On the other hand, if bands or band packets are sequentially or asynchronously recorded by different cameras, image cubes must be created by band coregistration. These camera systems are known as image frame cameras [85]. Fast cameras have significantly faster image acquisition speeds than those of image frame cameras, but faster speed also results in lower n class="Disease">hyperspectral image spatial resolutionpan>. However, the fast recording time anpan>d the conpan>tinuous creationpan> of n class="Disease">hyperspectral image sequences make the fast camera highly suitable for early detection of pine wilt disease in selected special environments. These nonintrusive sensors can complete image evaluation and analysis within a short time and have high image spatial resolution [86]. The combination of fast cameras and drones can collect a high amount of data and improve the efficiency of information acquisition [87].

7.3. Hyperspectral Image Acquisition Environment

Hyperspectral images of pine pan> class="Disease">wood nematode disease are currently obtained in the laboratory or in the field under relatively stable conditions, but research objects are mostly located in high mountains, steep roads, and dense forest areas, where ground investigation makes timely detection difficult, and environmental conditions are quite complicated.

Image acquisition in the field is affected by many factors, and the most important factor is light intensity [88]. Studies have shown that sunlight is an excellent light source for hyperspectral imaging, anpan>d covers a wide ranpan>ge of wavelengths. However, the main problem is that the light intensity of sunlight canpan>not be conpan>trolled. Usually, clouds cover the sun to varying degrees, causing uncertainty in light intensity. A chanpan>ge in light intensity during the time between measurement of the reference template anpan>d the sample canpan> lead to inaccurate measurements [89,90]. Light intensity is affected, anpan>d the anpan>gle of incident light chanpan>ges at each moment. The anpan>gle between incident light, planpan>t, anpan>d hyperspectral sensor has a certain effect on the spectral image acquisition of the sample plant [90,91,92,93]. This inaccurate information is gradually enlarged in subsequent information processing, which makes it difficult to obtain real information. Therefore, it is challenging to select light sources in the laboratory. Common light sources have the disadvantage of thermal changes. Generally, this problem can be solved by preheating [94]. To supply sufficient lighting conditions for hyperspectral imaging, many companies and researchers have offered multiple solutions through the implementation of design modifications. The light source commonly used in these solutions is a tungsten halide lamp with different improvements [93,95], and the emitted light is in a spectrum of 350–2500 nm [96]. Nagasubramanian and others selected two 70 W quartz–tungsten halide lamps to supply stable illumination in the study of soybean anthracnose [97]. The lamp was aimed at the sample stage at a 45° angle and at a distance of 54 cm. After data preprocessing, a 3D (3 dimensions), deep convolutional neural network (DCNN) model was used in classification. The model achieved a classification accuracy of 95.73%. In 2015, light-emitting diodes (LEDs) with an emission peak at 470 nm were used by Mahlein and colleagues to establish an illumination array. This array could improve the sensitivity of a hyperspectral imager in the blue spectrum [98].

7.4. Hyperspectral Data Analysis

7.4.1. Data Processing

After hyperspectral data were set, the white anpan>d black frames were acquired. In the laboratory, with the exceptionpan> of the computer, all other equipment was placed in a black box to eliminate the interference of external light. The white frame was obtained by measuring white reference objects with reflectanpan>ce close to 100%, such as pan> class="Chemical">barium sulfate, magnesium oxide, and white reference tiles. Black frames were obtained by closing the camera cover to obtain 0% reflectivity [99,100]. Black and white frames can be collected in the field by similar means, and a unified standard hyperspectral image is obtained after calibration. Even a hyperspectral image obtained in the laboratory still contains a large amount of noise, and a large signal-to-noise ratio affects the application of the data. To solve this problem, many studies have measured the sample multiple times during the data measurement stage [101], and used the measurements to reduce the effect of noise. To solve this problem in the data analysis stage, researchers preprocess the image via such methods as Savitzky–Golay (SG) smoothing, standard normalisation (SNV), de-enveloping (CR), logarithmic inverse (LR), normalisation (NDV), and first and second derivatives [92,102,103,104]. The preprocessing of hyperspectral images improves the signal-to-noise ratio of the data, and makes useful information more prominent in the image. In addition, reducing the dimension and denoising the hyperspectral images are more conducive to building accurate models. However, the choice between the preprocessing and machine learning methods has an impact on model construction. Liu and others tested a variety of preprocessing techniques, and used two different machine learning methods. The preprocessing results via SG smoothing combined with partial least-squares regression (PLS) were considered the best, which predicted best for the content of N and P in citrus leaves [102].

7.4.2. Machine Learning Methods Used in PWN Research

In recent years, the application of hyperspectral technology has ranpan>ged from the remote sensing of the Earth’s surface to the identificationpan> of all aspects of microbial diseases in planpan>t pathology [105], leading to anpan> increase in the number of different data anpan>alysis methods that have been developed. This sectionpan> describes the machine learning methods used in pan> class="Species">PWN research. Figure 4 shows the construction process of PWN grading models. Zhang used a spectral angle mapper (SAM) to classify hyperspectral images obtained in field remote sensing research on nematode disease [106]. Spectral angle mappers can reduce the impact of light and terrain, and the classification threshold is lower. The obtained results are more authentic [107,108], and the lower the classification threshold of the SAM, the higher the accuracy. In this study, when the system threshold was set to 0.1, the nematode thresholds in the late and middle stages were 0.06 and 0.01, respectively, and high reliability was obtained in large-scale remote sensing. Lu performed artificial neural network (ANN) research on the prediction of PWD [109]. Artificial neural network regression is a method that uses neural networks to propagate between layers, simulating human brain receiver and message-processing activity [110,111]. ANN is also a learning classification method based on large samples. However, with this method, it is easy to overlearn and reduce the generalisation ability due to the influence of network structure and sample complexity. The most important parameter in the neural network regression model is the number of neurons [112]. The application of ANN is highly extensive, involving all aspects of scientific production [113]. Six stages were adopted in this research to differentiate healthy from diseased pines: healthy, infected, early stage of disease, middle stage of disease, serious stage of disease, and terminal stage of disease [109,114]. Three neural networks were used in the analysis of these six stages. The average accuracy rates of prediction were 64.0% for the BP (back propagation) neural network [115], 72.0% for radial basis neural network (RBF) [116], and 72.3% for the Elman neural network (REN) [117]. Liu of Nanjing Forestry University used a variety of machine learning methods to analyse data in the study of pine physiological parameters under the stress of pine wood nematodes [118]; these methods included support vector machine (SVM) [119], random forest (RF) [120], and the BP neural network. For the analysis by these methods, the R2 value of the chlorophyll content estimation model of high-intensity Pinus was 0.8234 using SVM, 0.7214 using RF, and 0.9022 using the BP neural network. Hence, artificial neural networks showed the best prediction performance.

Figure 4

Construction of PWN grading models. PWNL: pine wood nematode, Bursaphelenchus xylophilus.

7.5. Challenges and Countermeasures in PWD Monitoring with Spectral Technology

7.5.1. Challenges

(1) Problems with data acquisition equipment. Currently, more effective early detection is needed, and early warnings that can be observed before n class="Gene">PWD are required. Regular inspectionpan>s are required, but pan> class="Species">humans cannot do this work. The low endurance of machines and the low resolution of their sensors are not sufficient for solving this problem. However, with the introduction of more compact high-resolution sensors, this problem can be solved. Data collected by a high-resolution drone can be verified and classified in the original model to obtain an early warning, or the drone can quickly obtain its own analysis results for classification and marking in a self-contained microcomputer.

(2) Data acquisition environment problems. Sunlight, wind speed, humidity, temperature, and other factors can cause changes in the n class="Disease">hyperspectral image of the sample. When using anpan> pan> class="Chemical">UAV in the field for cruise warning activities, it is necessary to choose conditions similar to those in the laboratory. The goal is to make it easier to translate scientific results into PWD field monitoring.

(3) Construction of a data analysis prediction model. This is a substantial problem in the early monitoring of n class="Gene">PWD anpan>d the n class="Disease">hyperspectral monitoring of many diseases. Model construction is carried out with the support of data, which means that additional data are required for analysis, such that the model can obtain higher accuracy. Stable data collection must be performed under laboratory conditions, and data collected in the field need additional data corrections, such as for light radiation and ground objects.

(4) Time gap in early detection. There may exist a short time gap between detecting time and n class="Gene">PWD revealed time when we try early detection. It canpan> be one- or two-week difference, so adequate isolation or cutting canpan> be the manpan>agement option at the current stage.

7.5.2. Conclusions and Suggestions

In the current research, multispectral imagers in combination with n class="Chemical">UAV technology have mostly been used to directly acquire spectral images in the field. On the onpan>e hanpan>d, multispectral imagers have been miniaturised, anpan>d their weight anpan>d volume are compatible for their use with dronpan>es. On the other hanpan>d, manpan>y studies are still in the initial stages, with the conpan>cept of digital phenotype still being tested, anpan>d fewer studies are available regarding the spectral characteristics of the early symptoms of the disease. The early detectionpan> of disease canpan> provide stronpan>g evidence for the conpan>trol of the source of tranpan>smissionpan>, but internal chanpan>ges of the host during the incubationpan> period or in the early stage of the disease are indistinguishable to the naked eye; the high resolutionpan> of hyperspectral imagers can offer more fine-grained data for analysis. Currently, the portability of hyperspectral imagers offers greater possibilities for effective application of this technology. In terms of stability, the distance between the camera and target object severely affects data stability. Distance control is particularly critical for dynamic acquisition in the field. Distance affects the exposure of natural light and the camera’s range of acquisition. Introducing a wide-angle or fisheye lens to increase the acquisition range and reduce the number of acquisitions could greatly reduce image instability. It should be noted that, in order to reduce the influence of the external environment and ensure the accuracy of experimental results, the hyperspectral camera should be used at noon on sunny day. Owing to technical limitations, the signal-to-noise ratio of the obtained spectral data is too low. The mainstream method uses a variety of preprocessing methods and useful data learning models to analyse data. After the pine is infected with PWD, the pine needles begin to turn yellow, and the difference of hyperspectral reflectivity becomes more obvious. Therefore, it is much easier to monitor pine wilt disease using hyperspectral images, but the prediction models need to be retrained using samples infected with PWD. With the continued development of science, the technology of hyperspectral monitoring has the potential to make greater contributions to the monitoring of PWD.