Next Article in Journal
MD3: Model-Driven Deep Remotely Sensed Image Denoising
Previous Article in Journal
D3CNNs: Dual Denoiser Driven Convolutional Neural Networks for Mixed Noise Removal in Remotely Sensed Images
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 15
Issue 2
10.3390/rs15020444
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Detection of the Monitoring Window for Pine Wilt Disease Using Multi-Temporal UAV-Based Multispectral Imagery and Machine Learning Algorithms

by
Dewei Wu
1,
Linfeng Yu
1,2,
Run Yu
1,
Quan Zhou
1,
Jiaxing Li
1,
Xudong Zhang
1,
Lili Ren
1,3 and
Youqing Luo
1,3,*
1
Beijing Key Laboratory for Forest Pest Control, Beijing Forestry University, Beijing 100083, China
2
School of Ecology and Nature Conservation, Beijing Forestry University, Beijing 100083, China
3
Sino-French Joint Laboratory for Invasive Forest Pests in Eurasia, Beijing Forestry University—French National Research Institute for Agriculture, Food and Environment (INRAE), Beijing 100083, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2023, 15(2), 444; https://doi.org/10.3390/rs15020444
Submission received: 5 December 2022 / Revised: 8 January 2023 / Accepted: 9 January 2023 / Published: 11 January 2023
Browse Figures
Versions Notes

Abstract

:
Pine wilt disease (PWD) has caused huge economic and environmental losses since it invaded China. Although early monitoring is an effective way to control this hazard, the monitoring window for the early stage is hard to identify, and varies in different hosts and environments. We used UAV-based multispectral images of Pinus thunbergii forest in East China to identify the change in the number of infected trees in each month of the growing season. We built classification models to detect different PWD infection stages by testing three machine learning algorithms—random forest, support vector machine, and linear discriminant analysis—and identified the best monitoring period for each infection stage (namely, green attack, early, middle, and late). From the obtained results, the early monitoring window period was determined to be in late July, whereas the monitoring window for middle and late PWD stages ranged from mid-August to early September. We also identified four important vegetation indices to monitor each infection stage. In conclusion, this study demonstrated the effectiveness of using machine learning algorithms to analyze multitemporal multispectral data to establish a window for early monitoring of pine wilt disease infestation. The results could provide a reference for future research and guidance for the control of pine wilt disease.

1. Introduction

Pine wilt disease is caused by the pinewood nematode Bursaphelenchus xylophilus (Steiner & Buhrer) Nickle, which is mainly transmitted by insects of the genus Monochamus [1,2,3,4]. The main transmission vectors in China are Monochamus alternatus (Hope) and Monochamus saltuarius (Gebler) [5,6]. Vector insects spread pinewood nematodes to the host tree when they feed on branches and lay eggs in tree trunks, resulting in the death of the trees. B. xylophilus was first discovered in China in 1982 on the black pine of the Sun Yat-sen Mausoleum in Nanjing [7]. In the last 40 years, pine wilt disease has gradually expanded nationwide and killed billions of pine trees, causing direct and indirect economic losses of over 100 billion yuan [8]. At present, there are 17 species of pine trees that are naturally infected by pine wilt disease in China [7,9,10], and four species are found in Shandong epidemic areas: Pinus thunbergia (Parl.), P. densiflora (Sieb. et Zucc.), P. massoniana (Lamb.) and P. tabuliformis (Carr.) [11]. According to the latest announcement of the State Forestry and Grassland Administration, pine wilt disease in China in 2021 is spread over 721 epidemic areas in 17 provinces (autonomous regions and municipalities) (State Forestry and Grassland Administration Announcement, 2021 no. 5). It can be considered one of the most serious and costly forest-related disasters recorded in China.
At present, the most commonly adopted control measure is the logging of infected trees [12,13]. Accurate monitoring is the basis for the effective felling of diseased trees [14]. However, traditional ground surveys are time-consuming and labor-intensive. High-resolution remote sensing can effectively solve this disadvantage. The most significant visible changes in pine trees infected with pinewood nematodes include two features: the discoloration of pine needles, and the reduction in pine resin secretion. Both of them are due to the changes in physiological parameters (such as photosynthesis and water content) and biochemical parameters (such as chlorophyll content) inside the pine trees [15]. Compared with needles infected by PWD, healthy needles exhibit higher reflectance in blue and green bands and lower reflectance in the red band [16]. Under the stress of PWD, plants will block the water metabolism of leaf cells, destroy the cell structure and reduce the content of internal pigments. These changes will lead to the decrease of absorption in the red band and the increase in spectral reflectance [17,18,19]. These physiological and biochemical transformations can be detected through changes in the spectral values of remote sensing images, making it possible for remote sensing monitoring of susceptible pine trees at early stages [20].
In recent years, research on monitoring pine wilt disease using unmanned autonomous vehicles (UAV) as data sources has been very extensive, and can be roughly divided into the following aspects. To begin with, there was the monitoring of diseased trees at the late stage of pine wilt disease based on RGB images and multispectral images, primarily focusing on the identification and classification of dead trees [21,22,23]. Li et al. [24] used a UAV with an RGB camera and an image segmentation algorithm based on the combination of the ultra-green feature factor and the maximum interclass variance method (ExG + Otsu) to identify dead trees with an accuracy of 90.4%. Zhang et al. [25] used UAV aerial photography to acquire RGB images of infected pine forests over large areas, and combined them with the deep learning segmentation network U-Net to carry out the image segmentation of infected pine trees with a recognition accuracy of 95.17%. However, because visible light reflectance is generally not suitable for the early monitoring of infected trees, there is little research in this area.
Hyperspectral UAV data have higher spectral resolution and are often used in high-precision early monitoring [26,27]. A few studies have used hyperspectral data for early monitoring of pine wilt disease [28,29]. Yu et al. [30] proposed a new approach combining the metrics of UAV-based hyperspectral imaging and light detection and ranging (LiDAR) data to precisely predict pine wilt disease infection stages at the tree level, obtaining good accuracy (overall accuracy (OA) 73.96%, kappa 0.66). However, hyperspectral image processing is complex and expensive and is not suitable for large-scale applications. There have also been relevant studies on the early detection of pine wilt disease using UAV multispectral data. Yu et al. [16] used multispectral UAV data of infected pine forests and employed two object detection algorithms (Faster R-CNN and YOLOv4) and two traditional feature extraction-based machine learning algorithms (random forest, RF, and support vector machine, SVM) to identify infected pines with an accuracy of 60.98–66.7%. The current early monitoring was aimed at identifying infected trees when their canopy has already turned yellow-green, but this was not the real green attack stage. The infected trees with green crowns were defined as trees in the green attack stage, which need to be carefully monitored. In early monitoring, most studies analyzed single images rather than multitemporal data [31].
According to the previous research studies, we believe that the use of multispectral images to analyze the green attack stage will be helpful for the early monitoring of pinewood nematode disease. In this study, we used a UAV with multispectral sensors to capture multitemporal images to observe the spectral changes of infected trees damaged by pinewood nematodes. This paper aimed to: (1) quantify changes of infected trees in different months of one infection cycle; (2) build classification models of healthy and diseased trees by analyzing multitemporal multispectral data with three machine learning algorithms—random forest (RF), support vector machine (SVM), and linear discriminant analysis (LDA); and (3) explore the monitoring window period of different infection stages, especially the early monitoring window. The results documented in this article provide technical support for the comprehensive control of pine wilt disease.

2. Materials and Methods

2.1. Study Area

A study area (Figure 1) of 6.52 ha was selected in Muping District, Yantai City, Shandong Province, East China (37.4499°N, 121.6997°E). The area has a warm, temperate, continental monsoon climate, with an annual average temperature of 13.2 °C, annual average precipitation of 57.25 mm, and annual average humidity of 64.58%. The tree species in the study area was the Japanese black pine (P. thunbergii). Japanese black pine is the main species of Yantai coastal protection forest, which has the functions of windbreak and sand fixation, water and soil conservation, and water conservation. It plays a unique and irreplaceable role in disaster prevention, mitigation and maintenance of ecological balance in coastal areas. According to the relevant forestry departments, the area was determined to be a pine wilt disease endemic area.

2.2. Field Measurement Data

The ground survey was carried out from May to November 2021. Ground surveys were conducted once a month. We established three plots of 30 × 30 m in the study area and inspected 15–20 pine trees in each plot. We selected some trees with slight discoloration from May to July to detect pinewood nematodes, mainly including damaged branches and wood cores. We used the Bellman funnel method [32] to obtain nematode extracts and used morphological and molecular identification techniques to detect whether the sampled trees were infected with pine wilt disease. At first, we obtained a low detection rate (3.92–25.33%). Therefore, we continued to sample and apply the same detection method once the tree crowns turned yellow-green and achieved a detection rate of 88.3%, which showed that these sample trees were infected with pine wilt disease. As a result, we found that at the early stage of infection, that is, when there was little difference between the tree crowns of diseased and healthy trees, it was difficult to detect pinewood nematodes due to their scarcity.

2.3. UAV-Based Multispectral Data Acquisition and Preprocessing

This study used the DJI Phantom 4 Multispectral Edition (Figure 2) to acquire images for the study and used virtual RTK (real-time kinematic) services for precise positioning. The device was equipped with a visible light lens and five multispectral lenses, including five bands, namely, blue (450 nm ± 16 nm), green (560 nm ± 16 nm), red (650 nm ± 16 nm), red edge (730 nm ± 16 nm), and near-infrared (840 nm ± 26 nm), and collected both visible light and multispectral images. The sensor has a field of view of 62.7° and a focal length of 5.74 mm. From 20 June to 24 November, the frequency of UAV image data acquisition was once every 3 days. The UAV images were taken in clear weather from 9:00 to 14:00. The flight altitude and speed were 30 m and 3 m/s, respectively, and the ground sample distance (GSD) was 1.59 cm/pixel. The front overlap and side overlap were set to 75%, and the flight cycle was 3 days. The standard board is covered with PTFE (polytetrafluoroethylene) material, which is a Lambertian reflector, for correction and calibration of the multispectral data. The reflectance of the standard board was known (0.6 in our study).
Due to the poor lighting conditions of some shooting dates, this part of the observation data was removed, and only the data of the remaining 10 observation dates were analyzed. Therefore, we selected multispectral images for 10 dates to build a multitemporal data sequence for the analysis. Terra v3.5.5 was used to perform atmospheric correction, radiation correction and stitching of UAV images. The final output was an orthophoto mosaic. ENVI v5.3 was used to analyze the stitched images in the next step. Regions of interest (ROIs) were traced to extract the canopy of sample trees in the superimposed stitched images. The ROI of each tree is manually extracted, and each ROI has more than 1000 pixels.
We selected a total of 246 Japanese black pine trees, of which 123 were healthy and 123 were infected. Previous studies mainly divided PWD infection stage into four classes: healthy, early stage, middle stage, and late stage [16]. We added a green attack stage between the healthy and early stages by combining multitemporal UAV multispectral images. This experiment was mainly based on the visual interpretation of multitemporal UAV images to determine the green attack stage. Firstly, we determined the UAV image of the sample tree in its early stage (the tree crown turns yellow-green for the first time, such as 30 July). Secondly, we selected the UAV images of the same sample tree in the green attack stage (the tree crown is green) for tree crown extraction in the UAV images from 20 to 29 July (within 10 days). The green attack stage was defined as the infected tree in the first 10 days before the early stage of infection (yellow-green), and was characterized by the canopy still being green. Pinewood nematodes were detected in all infected trees at the middle stage. Detailed tree apparent characteristics and classification criteria are shown in Table 1. Temporal changes in canopy color are shown in Figure 3.

2.4. Vegetation Index Selection

Referring to previous spectral studies on discolored trees and considering the changes in chlorophyll, water content and other aspects after tree damage [12,16,33], we analyzed 23 parameters: 5 individual bands and 18 vegetation indices. Specific parameters are shown in Table 2.
UAV data were collected from June to November. In order to exclude changes in target spectral indices not caused by pine wilt disease, such as changes caused by physiological reactions in the growing season, climate events, and atmospheric effects, and accurately obtain changes in spectral characteristics caused by pine wilt disease, this experiment referred to previous studies [52,53] to detrend the spectral data and reduce spectral differences caused by factors other than pine wilt disease. The formula used is as follows [53]:
Si = (S − Sμ)/Sδ
where S is the original spectral index value, Si is the rescaled spectral index, and Sμ and Sδ are the mean and standard deviation values, respectively, for the healthy trees within the same time step. Using this method, the time series of the spectral indices can be detrended in a robust and easy-to-implement manner, provided that the healthy trees selected were representative of the average forest conditions within a time step.
We further explored the impact of multitemporal differences in spectral characteristic data on model accuracy by calculating all parameters in the following way [14]:
Ci = Ai − Oi
where Ci represents the spectral difference between different dates, Ai represents the detrended spectral parameter value, and Oi represents the detrended initial value of each spectral parameter (the spectral parameter values of the same tree on 20 June were taken as the initial values because most trees in the study area were healthy on this date).

2.5. Data Analysis

The overall analysis process is shown in Figure 4.

2.5.1. Temporal Data Analysis

We explored the temporal relationship between the changes in tree spectral characteristics and disease progression of infected trees. We believed that the change in number of trees from the healthy stage to the infection stage was a good proxy of the dynamic processes of this disease in the study area. The cross-correlation factor (CCF) available from the “stats” R package was calculated between the temporal change in the number of healthy trees and the spectral indices examined. This analysis was designed to examine whether the spectral indices and the quantitative temporal change in healthy trees varied together or whether there was a temporal lag that would indicate that a spectral index was sensitive earlier or later to the temporal number change in healthy trees. Based on the statistical distribution of the temporal changes of parameters selected in the study, we set a certain range as “no-change” region. Our studies used the 25th and 75th percentiles of the indices extracted from healthy trees to define this region. Plots were considered stressed if the 75th percentile detrended spectral index values dropped below the lower bound of the no-change region or the 25th percentile detrended spectral index values rose above the upper bound of the no-change region.

2.5.2. Classification Procedure

We used RF, SVM and LDA for classifying the infected trees. RF is a nonparametric algorithm widely used in classification and regression. It is suitable for solving classification problems with a small number of samples and a large number of variables and is robust against multicollinearity and overfitting [54,55]. Two important parameters needed to be defined: the number of decision trees (ntree) and the number of variables/features at each split node (mtry) [56]. We randomly set relevant parameters and compared the model results to confirm the optimal parameter values. Therefore, we set the parameter “ntree” to 1000 and the parameter “mtry” to 3. The RF classification models were established using the statistical software R version 4.0.4 with the package “randomForest” [57,58].
SVM is another supervised machine learning algorithm applied for classification analysis. It has strong generalization ability and can simultaneously minimize empirical errors and maximize the geometric edge region [59]. It builds optimal hyperplanes in high-dimensional space based on the theory of minimizing structural risks. These hyperplanes increase the distance (margin) between the closest points in the two categories as much as possible, and the points lying on the boundaries, called support vectors, determine the margin [58,59,60,61]. The SVM type used in the experiment is C-classification, and the kernel is radial basis [62]. For SVM, the regularization or complexity parameter (C) and radial kernel search parameter (gamma) chosen to minimize the cross-validation error are default values. Relevant studies have proved that this method can effectively classify remote sensing data [54,63]. SVM classification models were established using R statistical software version 4.0.4 with the package “e1071.”
LDA is an effective subspace technique, as it optimizes the Fisher score [64,65]. In addition, it does not require the tuning of free parameters. These good capabilities have resulted in its extensive use and practical exploitation in remote sensing applications mainly focused on image classification and feature reduction. LDA has been used in geobotanical investigation, conifer species recognition, classification of tropical rainforest tree species, and identification of land-cover units in ecology, among others [66,67,68,69]. The class proportions for the training set were used as the prior probabilities of class membership in the study. LDA classification models were established using R statistical software version 4.0.4 with the package “MASS.”
We used RF to calculate the importance of each feature (i.e., the vegetation indices and bands used for the classifications), and took MDA (mean decrease accuracy) under each observation date as the evaluation basis of the importance of each parameter in the classification accuracy of the infected trees.
We used the weighted overall accuracy (WO) value to explore the temporal change in classification accuracy. At the same time, we used Student’s t-test to analyze the significance of differences in the precision of the models fitted by the same algorithm, but with different data at each date. We used 60% of the data as the training set and 40% as the validation set, according to the requirements. The training set includes canopy spectral values and vegetation indices of 114 trees on 10 dates, and the validation set includes canopy spectral values and vegetation indices of 76 trees on 10 dates. The same algorithm was repeated three times, and the samples were randomly selected each time. The final result was a combination of the three results. The number of trees in each infection stage varied considerably on different observation dates. Therefore, we chose weighted precision to evaluate the total precision of the algorithm at different infection stages. The weighted overall accuracy was calculated as follows [70,71]:
WO = ∑ (UAi×Ni/N)
where UAi is the UA (user’s accuracy) of stage i, Ni is the number of stage i, and N is the total number of samples in the same observation day.
With regard to the weighted accuracy, after fully considering the number of samples, we set two thresholds for WO: 0.7, indicating relatively high accuracy, and 0.8, indicating high precision.

2.5.3. Definition of the Monitoring Window Period

The main goal of this study was to define the monitoring window period for pine wilt disease in the study area. We defined the monitoring window as the key observation period in the entire observation period during which a tree goes from healthy to dead. This period needs to meet two conditions: high precision and high proportion of infected trees. In this experiment, we assumed that high precision was 0.7 and above. Considering the demand of the algorithm for the number of samples, a high proportion was defined to be 20% and above.

3. Results

3.1. Monitoring Window Period of Infected Trees

3.1.1. Temporal Changes of the Number of Infected Trees

Combined with the multitemporal UAV image analysis, the time sequence changes of the number of five different infection stages of 123 infected trees are shown in Figure 5. From late June to mid-August, the number of trees in the healthy stage gradually decreased to fewer than 20. The results showed that there were two growth peaks in the number of trees at the early infection stage (late July and late August), corresponding to the two emergence peaks of M. alternatus in late June and late July. In the early stage, the number of infected trees gradually increased from the beginning, reached a peak around the first 10 days of August, and then gradually decreased until the end of August. There was a slight recovery, and then it gradually decreased. The number of infected trees in the middle stage gradually increased from the beginning, reached a peak in late August, and then gradually decreased. The number of trees in the late stage had been rising continuously since late July, and finally, more than 97.6% had changed to the late stage.

3.1.2. Monitoring Accuracy of Infected Trees

The WO values of the six curves increased gradually with the change in observation time (Figure 6). In late July, the precision of each model was as follows—RF: 0.645; RF-C: 0.675; SVM: 0.684; SVM-C: 0.693; LDA: 0.719; LDA-C: 0.680. On 10 August, the precision of each model was as follows—RF: 0.792; RF-C: 0.877; SVM: 0.829; SVM-C: 0.846; LDA: 0.864; LDA-C: 0.886. In the last 10 days of September, the precision of each model was as follows—RF: 0.999; RF-C: 1; SVM: 0.996; SVM-C: 1; LDA: 0.996; LDA-C: 0.991. It could be seen that the accuracy of the model calculated by using the Ci data had improved for most observation dates compared with the accuracy of the model calculated by using the Ai data. Taking the RF model as an example, from 23 July to 2 September, the accuracy of the model fitted with the Ci data was significantly improved compared with the model fitted with the Ai data.

3.1.3. Multitemporal Change in Important Classification Parameters

In this study, we focused on the period when the number of trees in the green attack stage (23 July) and the middle stage (24 August) significantly increased (Figure 7). On 23 July, NDRE, LCI, red edge, NIR and red had the most significant impact on accuracy. On 10 August, PSRI, green NDVI, CIG, PBI and NDVI had the most significant impact on accuracy.
We compared the original and the detrended values of each parameter of healthy and infected samples (Figure A1 of Appendix A). There were subtle multitemporal changes in the original values of an infected tree, but after detrending, there were obvious downward or upward trends. The multitemporal changes of LCI, Red, PSRI, NDVI, NWI, VARI and RVI were more significant. From 10 August, red and PSRI increased above the no-change region. Meanwhile, LCI, NDVI, NWI, VARI and RVI fell below the no-change region.
The multitemporal change in the red spectral characteristics of infected trees was particularly evident. Therefore, we used CCF to analyze the correlation between the quantitative temporal change in healthy trees and the temporal change in the red spectral values. The correlations were largest when the lag was equal to zero. This indicated that while the spectral indices were accurate indicators of the infection, they did not provide an earlier indication of stress compared with the UAV observations.

3.2. Multitemporal Changes of Parameters of Trees at Different Infection Stages

3.2.1. Spectral Characteristics of Trees at Different Infection Stages

Considering that the spectral characteristics of the targets in images of different observation dates differed slightly, we chose trees in different stages of the same day to compare their spectral characteristics. The results in Figure 8 show variations in the spectral characteristics of the red, red edge and NIR bands of trees at different infection stages. For example, the red band reflectance in the middle and the late stages was higher than those of the healthy, green attack, and early stages. The red edge and NIR reflectance of the stages with the greatest crown discoloration were lower than those of the stage with no or no significant discoloration of the crown.

3.2.2. Classification Accuracy at Different Stages of Infection

The results in Section 3.2.1 demonstrated that the accuracy obtained by fitting the model with Ci is higher than that obtained by fitting the model with Ai. Accuracy of the three algorithms increased as the infection stage progressed (Figure 9). The overall trend was the same as the results in Section 3.1.2. Before 17 August, the discrimination accuracy based on the three algorithms was lower than 0.7 and exceeded 0.8 after 2 September.
Accuracy of the models fitted by each algorithm on the same date showed that there was no significant difference in the precision of the models fitted by RF and SVM on the same observation date (Table 3). We selected the RF model and drew a confusion matrix for several key dates with improved accuracy (Table A1). The classification accuracy of each infection stage under different dates is shown in Table 4. The results showed that the WO and KAPPA continued to increase and the accuracy of each infection stage continued to increase. The number of samples at each stage also increased progressively; thus, the accuracy of each stage increased.

3.2.3. Important Parameters Affecting Monitoring Accuracy at Different Stages of Infection

The early monitoring window period and the middle and late infection window period were determined according to the multitemporal changes in the number of infected trees in different stages and the RF accuracy. The images of July and August were selected. The ranking results of parameter importance under each observation date are shown in Figure 10. On 23 July, the top five descending MDA parameters were NDRE, LCI, red, PSRI, and NWI. On 10 August, the top five descending MDA parameters were PSRI, NDVI, LCI, NWI, and green NDVI. On 24 August, the top five descending MDA parameters were PSRI, NDVI, NWI, VARI and NGRDI.
Due to the harm of pine wilt disease, the pigment and water content of the leaves decreased, and the spectral characteristics of the leaves changed. The reflectance of leaves in the red band increased, while the reflectance of the red edge and near-infrared bands decreased, and this change was relatively obvious in the early stage of infection. Therefore, the vegetation index constructed based on the reflectance of the above three bands was more sensitive to changes in physiological indicators than other vegetation indices, such as NDRE, LCI, PSRI, etc.

4. Discussion

Based on the multitemporal multispectral images and machine learning algorithms, this paper established the window periods of different infection stages of P. thunbergii along the coast of Yantai, Shandong with high monitoring accuracy. In comparison to previous studies, the results of this paper clarified that the classification accuracy of the model can be effectively improved by using multitemporal multispectral data.

4.1. Temporal Variation in Tree Numbers at Different Infection Stages

The results showed that there were two growth peaks in the number of trees at the early infection stage (late July and late August), corresponding to the two emergence peaks of M. alternatus in late June and late July. Based on this, we made it clear that the changes in the crown color of the trees in the study area can be observed with the naked eye about one month after being damaged by pine wilt disease. In this study, we also established a multitemporal change rule based on the number of infected trees in different infection stages, which allowed us to determine the monitoring window period of different infection stages based on the discrimination accuracy of different models. Multitemporal data are used to improve the monitoring accuracy in the early stages of disease, and there are few experiments involving the multitemporal changes of the number of infected trees [14,30,72].

4.2. Important Parameters of Different Monitoring Window Periods

The results showed that the effect of vegetation indices on the accuracy increased gradually with observation time. The MDA values of PSRI, NDRE, LCI and NDVI in different monitoring periods were all high, indicating that these indices played an important role in the monitoring of pine wilt at different infection stages. These indices are mostly based on the red, red edge and NIR bands. Our results were consistent with the conclusion that these bands and vegetation indices are important for the monitoring of pine wilt disease [12,26,73]. However, the CCF result indicated that the red band value provided an accurate indicator of UAV measured values, but did not provide an earlier indication of stress than the UAV observations.

4.3. Monitoring Window Period

In this paper, we built a window period for monitoring different stages of infected trees for different actual monitoring needs. In terms of early monitoring, we focused on comparing the differences between the healthy stage and the green attack stage. At the end of July, the WO of the three algorithms fitting the model exceeded 0.7, and the number of trees in the green attack infection stage was at a peak. Therefore, the early monitoring window period of pine wilt disease in the study area was late July. Because there are ancient trees in the study area, accurate monitoring of infected trees in the early monitoring window will help the relevant departments take relevant control measures as soon as possible.
On the other hand, it is very important to determine the monitoring window in the middle and late infection stages of pine forests in the epidemic areas where pine trees are seriously affected by pine wilt disease. From the middle of August to the beginning of September, there was a large proportion of trees in the middle infection period, and the WO of the models fitted by the three algorithms exceeded 0.8. Based on the analysis of the multitemporal data, we found that the increase in the number of samples in the middle and late infection stages was the main reason for the improvement in accuracy as the observation date progressed. Therefore, we believed that the monitoring window period for the middle and late stages of pine wilt disease was from mid-August to early September. The monitoring of infected trees in the middle and late stages of infection would help to count and locate the number of infected trees to be cleaned and take effective measures to prevent and control them.
In this paper, we tried to clarify the window period of different infection stages in the research area (Figure 11), but the determination of the monitoring window period of PWD in different regions needed to consider more factors. When it was extended to a wider area, factors such as environmental conditions in different areas, local tree species, and the life history of local media needed to be further considered in detail [16,29,72].

4.4. Advantages of Multitemporal Data

The results showed that detrending spectral data could not only effectively identify stressed trees but also effectively reduce the noise caused by weather or other factors. In addition, compared with the model calculated with Ai data, the accuracy of the model calculated with Ci data was improved on most observation dates. It can be seen that the time-series differences of parameters obtained by multitemporal observations can effectively improve the classification accuracy of the model. From 23 July to 2 September, the accuracy of the model fitted by RF was further improved by 3–8%. Operationally, relevant personnel should acquire multispectral UAV images according to the life history of the vector beetles. The images should be taken before peak emergence and one month and three months after peak emergence. The damage of pine wilt disease at different infection stages could be accurately monitored by fitting the model presented here.
Since the early monitoring of infection in this article mainly focused on the green attack stage, there were few discussions on this stage in existing articles. Therefore, the comparison of the early detection accuracy of infection mainly focused on the comparison of the early stage, because the monitoring accuracy for the PWD early stage on different dates continued to change, and the number of infected trees was the main reason for the change. Considering that in forestry work, when there was a certain number of infected trees, the work carried out can maximize the cost savings and achieve the expected results. Therefore, this paper compared the early detection accuracy on August 10, when the number of trees in the early stage of infection reached its peak, with the accuracy of early detection of PWD detection in other studies.
Compared with the early monitoring accuracy of infection obtained by using single-temporal multispectral data, this experiment has the following advantages. First, we have obtained a higher monitoring accuracy. The accuracy obtained by using the RF fitting model is 0.615, which is about 12.6% higher than the 0.489 obtained by using the Faster R-CNN fitting model in other experiments [16]. Second, we have made it clear that higher early detection accuracy can be obtained at a relatively early time. Other experiments using single-temporal data can only obtain monitoring accuracy for a specific day. Third, we can focus on monitoring the infected trees in the green attack stage.
Many studies on remote sensing of pine wilt disease have focused on using various algorithms to improve the classification accuracy of damaged wood (red crown) and dead wood (gray crown), but there are still few studies on the monitoring date [74,75,76,77]. Because of unsuitable weather conditions during the observation period, this experiment lacked more frequent data in June and July. Increasing the observation frequency and exploring the data potential under poor lighting conditions might help to narrow the time range of the monitoring window period. At the same time, using multitemporal hyperspectral data for analysis and selecting more algorithms, such as deep learning, could further improve the accuracy of the early monitoring window period.

5. Conclusions

In this paper, time series of multispectral UAV images were used to successfully monitor pine wilt disease in coastal black pine forests. First, in the healthy and early stages, the green attack stage was further proposed as the key stage of early monitoring. Second, multitemporal UAV images illustrated the temporal changes in the number of trees in each infection stage. Third, LDA, SVM and RF were used to fit the trend spectrum data, and the infection monitoring window period was defined. The early monitoring window period was the end of July, and the monitoring window period for the middle and late stages of pine wilt disease was from the middle of August to the beginning of September. Fourth, the order of different parameters under different monitoring purposes was clarified through RF. The results showed that PSRI, NDRE, LCI and NDVI were of great significance in monitoring different infection stages. NDRE is the best vegetation index for early detection. The results of this paper provide data support for precise monitoring and scientific prevention and control of pine wilt disease.
We believe that accurate monitoring is the basis for effective control of pine wilt disease. At present, we mainly focus on using traditional algorithms to analyze multitemporal and multispectral data, fully mining the positive impact of multitemporal data on classification accuracy, and reducing the difficulty of practical application. From the division of infection stage, we mainly monitor the green attack stage to achieve the purpose of early monitoring. In the future, we can pay more attention to new findings and new methods of multitemporal multispectral data processing and achieve higher monitoring accuracy. The establishment of monitoring window period is conducive to early detection of damaged trees, and helps forestry departments to take control measures, such as cutting down damaged trees. The fallen trees will be burned or crushed on the spot to kill nematodes and prevent large-scale forest destruction.

Author Contributions

Conceptualization, Y.L. and L.R.; methodology, D.W. and L.Y.; software, D.W.; validation, D.W.; formal analysis, D.W.; investigation, D.W., R.Y., Q.Z., J.L. and X.Z.; writing—original draft preparation, D.W.; writing—review and editing, D.W., L.Y. and L.R.; project administration, Y.L.; funding acquisition, Y.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Key R&D Program of China (2021YFD1400900) and Major Emergency Science and Technology Project of National Forestry and Grassland Administration (ZD202001).

Data Availability Statement

Not applicable.

Acknowledgments

The author would like to thank Yantai Forest Resources Service Center and Muping District Forest Defense Station for their strong support in the field survey. At the same time, the author would like to thank the students of Beijing Key Laboratory for Forest Pest Control for their strong support in data processing.

Conflicts of Interest

The authors declare no conflict of interest.

Appendix A

The timing changes of each parameter are shown in Figure A1 below.
Remotesensing 15 00444 g0a1a 550Remotesensing 15 00444 g0a1b 550Remotesensing 15 00444 g0a1c 550Remotesensing 15 00444 g0a1d 550Remotesensing 15 00444 g0a1e 550Remotesensing 15 00444 g0a1f 550
Figure A1. Multitemporal change in each parameter. Upper and lower red dotted lines indicate the extent of the “no-change” region. The no-change region was defined as the 25th and 75th percentiles of the spectral value prior to infection (20 June 2020). The black horizontal line is the median line; the diamond point an abnormal value; and the square point in the middle of the box is the mean value. The range of the box line is 1.5 IQR. (a) NDRE; (b) LCI; (c) red edge; (d) NIR; (e) red; (f) PSRI; (g) NWI; (h) VARI; (i) NDVI; (j) RVI.
Figure A1. Multitemporal change in each parameter. Upper and lower red dotted lines indicate the extent of the “no-change” region. The no-change region was defined as the 25th and 75th percentiles of the spectral value prior to infection (20 June 2020). The black horizontal line is the median line; the diamond point an abnormal value; and the square point in the middle of the box is the mean value. The range of the box line is 1.5 IQR. (a) NDRE; (b) LCI; (c) red edge; (d) NIR; (e) red; (f) PSRI; (g) NWI; (h) VARI; (i) NDVI; (j) RVI.
Remotesensing 15 00444 g0a1aRemotesensing 15 00444 g0a1bRemotesensing 15 00444 g0a1cRemotesensing 15 00444 g0a1dRemotesensing 15 00444 g0a1eRemotesensing 15 00444 g0a1f
Table
Table A1. Confusion matrix for several key days.
Table A1. Confusion matrix for several key days.
Date23 July
Reference dataClassified as
Infection stagesHGAEMLUA
H1130000.786
GA491000.643
E031000.250
M00000NA
L010000
PA0.7330.5630.500NANA
WO0.636
KAPPA0.388
Date10 August
Infection stagesHGAEMLUA
H1210000.923
GA112000.250
E118300.615
M003700.700
L000021.000
PA0.8570.3330.6150.7001.000
WO0.714
KAPPA0.611
Date24 August
Infection stagesHGAEMLUA
H1700001
GA202000
E205100.625
M0011420.824
L000350.625
PA0.81NA0.6250.7780.714
WO0.759
KAPPA0.671
Note: NA means not applicable.

References

  1. Kim, N.; Jeon, H.W.; Mannaa, M.; Jeong, S.I.; Kim, J.; Kim, J.; Lee, C.; Park, A.R.; Kim, J.C.; Seo, Y.S. Induction of resistance against pine wilt disease caused by Bursaphelenchus xylophilus using selected pine endophytic bacteria. Plant Pathol. 2019, 68, 434–444. [Google Scholar] [CrossRef]
  2. Hunt, D. Pine wilt disease: A worldwide threat to forest ecosystems. Nematology 2009, 11, 315–316. [Google Scholar] [CrossRef] [Green Version]
  3. Abelleira, A.; Picoaga, A.; Mansilla, J.P.; Aguin, O. Detection of Bursaphelenchus Xylophilus, Causal Agent of Pine Wilt Disease on Pinus pinaster in Northwestern Spain. Plant Dis. 2011, 95, 776. [Google Scholar] [CrossRef] [PubMed]
  4. Futai, K. Pine Wood Nematode, Bursaphelenchus xylophilus. Annu. Rev. Phytopathol. 2013, 51, 61–83. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  5. Zhang, J.; Zhang, R.; Chen, J. Species and dispersal ability of Bursaphelenchus xylophilus vector insects. J. Zhejiang For. Univ. 2007, 24, 7. [Google Scholar]
  6. Zhang, R.; You, J.; Lee, J. Detecting Pine Trees Damaged by Wilt Disease Using Deep Learning Techniques Applied to Multi-Spectral Images. IEEE Access 2022, 10, 39108–39118. [Google Scholar] [CrossRef]
  7. Zheng, Y.; Liu, P.; Shi, Y.; Wu, H.; Yu, H.; Jiang, S. Difference analysis of pine wilt disease in Liaoning and other endemic areas in China. J. Beijing For. Univ. 2021, 43, 155–160. [Google Scholar]
  8. Ye, J. Analysis on the Epidemic Status, Control Techniques and Countermeasures of Pine Wood Nematode Disease in China. For. Sci. 2019, 55, 1–10. [Google Scholar]
  9. Yu, H.; Wu, H.; Huang, R.; Wang, J.; Zhang, R.; Song, Y. Isolation and identification of Bursaphelenchus xylophilus in Fushun, Liaoning. China For. Pests 2020, 39, 6–10. [Google Scholar]
  10. Pan, L.; Li, Y.; Liu, Z.; Meng, F.; Chen, J.; Zhang, X. Isolation and identification of Bursaphelenchus xylophilus from Korean pine in Fengcheng City, Liaoning. China For. Pests 2019, 38, 1–4. [Google Scholar]
  11. Li, Y.; Zhang, X. Trend analysis of the invasion and expansion of Bursaphelenchus xylophilus. China For. Pests 2018, 37, 1–4. [Google Scholar]
  12. Kim, S.; Lee, W.; Lim, C.; Kim, M.; Kafatos, M.C.; Lee, S.; Lee, S. Hyperspectral analysis of pine wilt disease to determine an optimal detection index. Forests 2018, 9, 115. [Google Scholar] [CrossRef] [Green Version]
  13. Syifa, M.; Park, S.; Lee, C. Detection of the pine wilt disease tree candidates for drone remote sensing using artificial intelligence techniques. Engineering 2020, 6, 919–926. [Google Scholar] [CrossRef]
  14. Zhang, B.; Ye, H.; Lu, W.; Huang, W.; Wu, B.; Hao, Z.; Sun, H. A Spatiotemporal Change Detection Method for Monitoring Pine Wilt Disease in a Complex Landscape Using High-Resolution Remote Sensing Imagery. Remote Sens. 2021, 13, 2083. [Google Scholar] [CrossRef]
  15. Xu, H. Pathophysiology of Black pine and Masson pine after Natural Infection with Pine xylophilus. Diploma Thesis, Beijing Forestry University, Beijing, China, 2013. [Google Scholar]
  16. Yu, R.; Luo, Y.; Zhou, Q.; Zhang, X.; Wu, D.; Ren, L. Early detection of pine wilt disease using deep learning algorithms and UAV-based multispectral imagery. For. Ecol. Manag. 2021, 497, 119493. [Google Scholar] [CrossRef]
  17. Thomas, J.R.; Namken, L.N.; Oerther, G.F.; Brown, R.G. Estimating leaf water content by reflectance measurements 1. Agron. J. 1971, 63, 845–847. [Google Scholar] [CrossRef]
  18. Tao, H.; Li, C.; Cheng, C.; Jiang, L.; Hu, H. Research progress in remote sensing monitoring of pine wood nematode disease color. For. Sci. Res. 2020, 33, 172–183. [Google Scholar]
  19. Dawson, T.P.; Curran, P.J. Technical note A new technique for interpolating the reflectance red edge position. Int. J. Remote Sens. 1998, 19, 2133–2139. [Google Scholar] [CrossRef]
  20. Proença, D.N.; Francisco, R.; Santos, C.V.; Lopes, A.; Fonseca, L.; Abrantes, I.M.; Morais, P.V. Diversity of bacteria associated with Bursaphelenchus xylophilus and other nematodes isolated from Pinus pinaster trees with pine wilt disease. PLoS ONE 2010, 5, e15191. [Google Scholar] [CrossRef] [Green Version]
  21. Deng, X.; Tong, Z.; Lan, Y.; Huang, Z. Detection and Location of Dead Trees with Pine Wilt Disease Based on Deep Learning and UAV Remote Sensing. Agriengineering 2020, 2, 294–307. [Google Scholar] [CrossRef]
  22. Tao, H.; Li, C.; Zhao, D.; Deng, S.; Hu, H.; Xu, X.; Jing, W. Deep learning-based dead pine tree detection from unmanned aerial vehicle images. Int. J. Remote Sens. 2020, 41, 8238–8255. [Google Scholar] [CrossRef]
  23. You, J.; Zhang, R.; Lee, J. A Deep Learning-Based Generalized System for Detecting Pine Wilt Disease Using RGB-Based UAV Images. Remote Sens. 2021, 14, 150. [Google Scholar] [CrossRef]
  24. Li, H.; Xu, H.; Zheng, H.; Chen, X.Y. Research on pine wood nematode surveillance technology based on unmanned aerial vehicle remote sensing image. J. Chin. Agric. Mech. 2020, 41, 170–175. [Google Scholar]
  25. Zhang, R.; Xia, L.; Chen, L.; Xie, C.; Chen, M.; Wang, W. Recognition of wilt wood caused by pine wilt nematode based on U-Net network and unmanned aerial vehicle images. Trans. Chin. Soc. Agricult. Eng. 2020, 36, 61–68. [Google Scholar]
  26. Iordache, M.; Mantas, V.; Baltazar, E.; Pauly, K.; Lewyckyj, N. A machine learning approach to detecting pine wilt disease using airborne spectral imagery. Remote Sens. 2020, 12, 2280. [Google Scholar] [CrossRef]
  27. Einzmann, K.; Atzberger, C.; Pinnel, N.; Glas, C.; Boeck, S.; Seitz, R.; Immitzer, M. Early detection of spruce vitality loss with hyperspectral data: Results of an experimental study in Bavaria, Germany. Remote Sens. Environ. 2021, 266, 112676. [Google Scholar] [CrossRef]
  28. Wu, B.; Liang, A.; Zhang, H.; Zhu, T.; Zou, Z.; Yang, D.; Tang, W.; Li, J.; Su, J. Application of conventional UAV-based high-throughput object detection to the early diagnosis of pine wilt disease by deep learning. For. Ecol. Manag. 2021, 486, 118986. [Google Scholar] [CrossRef]
  29. Yu, R.; Luo, Y.; Li, H.; Yang, L.; Huang, H.; Yu, L.; Ren, L. Three-Dimensional Convolutional Neural Network Model for Early Detection of Pine Wilt Disease Using UAV-Based Hyperspectral Images. Remote Sens. 2021, 13, 4065. [Google Scholar] [CrossRef]
  30. Yu, R.; Luo, Y.; Zhou, Q.; Zhang, X.; Wu, D.; Ren, L. A machine learning algorithm to detect pine wilt disease using UAV-based hyperspectral imagery and LiDAR data at the tree level. Int. J. Appl. Earth Obs. Geoinf. 2021, 101, 102363. [Google Scholar] [CrossRef]
  31. Ye, H.; Huang, W.; Huang, S.; Cui, B.; Dong, Y.; Guo, A.; Ren, Y.; Jin, Y. Identification of banana fusarium wilt using supervised classification algorithms with UAV-based multi-spectral imagery. Int. J. Agric. Biol. Eng. 2020, 13, 136–142. [Google Scholar] [CrossRef]
  32. Song, Y.; Zang, X.; Liu, Y.; Wang, Y. Relationship between room temperature changes and pine xylophilus segregation. For. Dis. Pest Commun. 1992, 21–22. [Google Scholar]
  33. Zhou, Q.; Zhang, X.; Yu, L.; Ren, L.; Luo, Y. Combining WV-2 images and tree physiological factors to detect damage stages of Populus gansuensis by Asian longhorned beetle (Anoplophora glabripennis) at the tree level. For. Ecosyst. 2021, 8, 1–12. [Google Scholar] [CrossRef]
  34. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef] [Green Version]
  35. Haboudane, D.; Miller, J.R.; Pattey, E.; Zarco-Tejada, P.J.; Strachan, I.B. Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture. Remote Sens. Environ. 2004, 90, 337–352. [Google Scholar] [CrossRef]
  36. Gobron, N.; Pinty, B.; Verstraete, M.M.; Widlowski, J. Advanced vegetation indices optimized for up-coming sensors: Design, performance, and applications. IEEE Trans. Geosci. Remote Sens. 2000, 38, 2489–2505. [Google Scholar]
  37. Gitelson, A.A.; Vina, A.; Arkebauer, T.J.; Rundquist, D.C.; Keydan, G.; Leavitt, B. Remote estimation of leaf area index and green leaf biomass in maize canopies. Geophys. Res. Lett. 2003, 30, 5. [Google Scholar] [CrossRef] [Green Version]
  38. Vincini, M.; Frazzi, E.; D’Alessio, P. A broad-band leaf chlorophyll vegetation index at the canopy scale. Precis. Agric. 2008, 9, 303–319. [Google Scholar] [CrossRef]
  39. Hunt Jr, E.R.; Daughtry, C.; Eitel, J.U.; Long, D.S. Remote sensing leaf chlorophyll content using a visible band index. Agron. J. 2011, 103, 1090–1099. [Google Scholar] [CrossRef] [Green Version]
  40. Abdullah, H.; Skidmore, A.K.; Darvishzadeh, R.; Heurich, M. Timing of red-edge and shortwave infrared reflectance critical for early stress detection induced by bark beetle (Ips typographus, L.) attack. Int. J. Appl. Earth Obs. Geoinf. 2019, 82, 101900. [Google Scholar] [CrossRef]
  41. Gitelson, A.A.; Merzlyak, M.N.; Lichtenthaler, H.K. Detection of Red Edge Position and Chlorophyll Content by Reflectance Measurements Near 700 nm. J. Plant Physiol. 1996, 148, 501–508. [Google Scholar] [CrossRef]
  42. Harfi, T.; Tabatabai, S.J. Effect of Nitrogen Level on Growth, and Relationships between Petiole Nitrate Level, Leaf Chlorophyll Index, and Hypocotyl Nitrate Level of Radish. Isfahan Univ. Technol.-J. Crop Prod. Process. 2015, 4, 203–213. [Google Scholar]
  43. Suits, G.H. The calculation of the directional reflectance of a vegetative canopy. Remote Sens. Environ. 1971, 2, 117–125. [Google Scholar] [CrossRef]
  44. Liu, H.Q.; Huete, A. A feedback based modification of the NDVI to minimize canopy background and atmospheric noise. IEEE Trans. Geosci. Remote Sens. 1995, 33, 457–465. [Google Scholar] [CrossRef]
  45. Qiu, Y.; Zhou, J.; Chen, J.; Chen, X. Spatiotemporal fusion method to simultaneously generate full-length normalized difference vegetation index time series (SSFIT). Int. J. Appl. Earth Obs. Geoinf. 2021, 100, 102333. [Google Scholar] [CrossRef]
  46. Roujean, J.; Breon, F. Estimating PAR absorbed by vegetation from bidirectional reflectance measurements. Remote Sens. Environ. 1995, 51, 375–384. [Google Scholar] [CrossRef]
  47. Broge, N.H.; Leblanc, E. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sens. Environ. 2001, 76, 156–172. [Google Scholar] [CrossRef]
  48. Gitelson, A.A.; Kaufman, Y.J.; Stark, R.; Rundquist, D. Novel algorithms for remote estimation of vegetation fraction. Remote Sens. Environ. 2002, 80, 76–87. [Google Scholar] [CrossRef] [Green Version]
  49. Uto, K.; Takabayashi, Y.; Kosugi, Y.; Ogata, T. Hyperspectral analysis of Japanese Oak wilt to determine normalized wilt index. In Proceedings of the IGARSS 2008—2008 IEEE International Geoscience and Remote Sensing Symposium, Boston, MA, USA, 7–11 July 2008; Volume 2, p. 295. [Google Scholar]
  50. Sims, D.A.; Gamon, J.A. Relationships between leaf pigment content and spectral reflectance across a wide range of species, leaf structures and developmental stages. Remote Sens. Environ. 2002, 81, 337–354. [Google Scholar] [CrossRef]
  51. Waser, L.T.; Küchler, M.; Jütte, K.; Stampfer, T. Evaluating the potential of WorldView-2 data to classify tree species and different levels of ash mortality. Remote Sens. 2014, 6, 4515–4545. [Google Scholar] [CrossRef] [Green Version]
  52. Healey, S.P.; Cohen, W.B.; Zhiqiang, Y.; Krankina, O.N. Comparison of Tasseled Cap-based Landsat data structures for use in forest disturbance detection. Remote Sens. Environ. 2005, 97, 301–310. [Google Scholar] [CrossRef]
  53. Dash, J.P.; Watt, M.S.; Pearse, G.D.; Heaphy, M.; Dungey, H.S. Assessing very high resolution UAV imagery for monitoring forest health during a simulated disease outbreak. Isprs-J. Photogramm. Remote Sens. 2017, 131, 1–14. [Google Scholar] [CrossRef]
  54. Gao, B.; Yu, L.; Ren, L.; Zhan, Z.; Luo, Y. Early Detection of Dendroctonus valens Infestation with Machine Learning Algorithms Based on Hyperspectral Reflectance. Remote Sens. 2022, 14, 1373. [Google Scholar] [CrossRef]
  55. Wan, L.; Cen, H.; Zhu, J.; Zhang, J.; Zhu, Y.; Sun, D.; Du, X.; Zhai, L.; Weng, H.; Li, Y. Grain yield prediction of rice using multi-temporal UAV-based RGB and multispectral images and model transfer–a case study of small farmlands in the South of China. Agric. For. Meteorol. 2020, 291, 108096. [Google Scholar] [CrossRef]
  56. Han, L.; Yang, G.; Dai, H.; Xu, B.; Yang, H.; Feng, H.; Li, Z.; Yang, X. Modeling maize above-ground biomass based on machine learning approaches using UAV remote-sensing data. Plant Methods 2019, 15, 1–19. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  57. Mullen, K. Early Detection of Mountain Pine Beetle Damage in Ponderosa Pine forests of the Black Hills Using Hyperspectral and WorldView-2 Data. Master’s Thesis, Minnesota State University, Mankato, MI, USA, 2016. [Google Scholar]
  58. Gutierrez, D.D. Machine Learning and Data Science: An Introduction to Statistical Learning Methods with R.; Technics Publications: Bernards, NJ, USA, 2015. [Google Scholar]
  59. Sothe, C.; De Almeida, C.M.; Schimalski, M.B.; La Rosa, L.; Castro, J.; Feitosa, R.Q.; Dalponte, M.; Lima, C.L.; Liesenberg, V.; Miyoshi, G.T. Comparative performance of convolutional neural network, weighted and conventional support vector machine and random forest for classifying tree species using hyperspectral and photogrammetric data. Gisci. Remote Sens. 2020, 57, 369–394. [Google Scholar] [CrossRef]
  60. Lesmeister, C. Mastering Machine Learning with R; Packt Publishing Ltd.: Birmingham, UK, 2017. [Google Scholar]
  61. Rumpf, T.; Mahlein, A.; Steiner, U.; Oerke, E.; Dehne, H.; Plümer, L. Early detection and classification of plant diseases with support vector machines based on hyperspectral reflectance. Comput. Electron. Agric. 2010, 74, 91–99. [Google Scholar] [CrossRef]
  62. Oommen, T.; Misra, D.; Twarakavi, N.K.; Prakash, A.; Sahoo, B.; Bandopadhyay, S. An objective analysis of support vector machine based classification for remote sensing. Math Geosci. 2008, 40, 409–424. [Google Scholar] [CrossRef]
  63. Guo, Y.; Fu, Y.; Hao, F.; Zhang, X.; Wu, W.; Jin, X.; Bryant, C.R.; Senthilnath, J. Integrated phenology and climate in rice yields prediction using machine learning methods. Ecol. Indic. 2021, 120, 106935. [Google Scholar] [CrossRef]
  64. Fisher, R.A. The use of multiple measurements in taxonomic problems. Ann. Eugen. 1936, 7, 179–188. [Google Scholar] [CrossRef]
  65. Bandos, T.V.; Bruzzone, L.; Camps-Valls, G. Classification of hyperspectral images with regularized linear discriminant analysis. IEEE Trans. Geosci. Remote Sens. 2009, 47, 862–873. [Google Scholar] [CrossRef]
  66. Schwaller, M.R. A geobotanical investigation based on linear discriminant and profile analyses of airborne thematic mapper simulator data. Remote Sens. Environ. 1987, 23, 23–34. [Google Scholar] [CrossRef]
  67. Gong, P.; Pu, R.; Yu, B. Conifer species recognition: An exploratory analysis of in situ hyperspectral data. Remote Sens. Environ. 1997, 62, 189–200. [Google Scholar] [CrossRef]
  68. Clark, M.L.; Roberts, D.A.; Clark, D.B. Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sens. Environ. 2005, 96, 375–398. [Google Scholar] [CrossRef]
  69. Lobo, A. Image segmentation and discriminant analysis for the identification of land cover units in ecology. IEEE Trans. Geosci. Remote Sens. 1997, 35, 1136–1145. [Google Scholar] [CrossRef]
  70. Xu, Z. Research on Subtropical Forest Monitoring Method Based on UAV Remote Sensing and AI Algorithm. Diploma Thesis, Jiangxi Agricultural University, Nanchang, China, 2021. [Google Scholar]
  71. Xu, Z.; Guo, X.; Zhu, A.; He, X.; Zhao, X.; Han, Y.; Subedi, R. Using deep convolutional neural networks for image-based diagnosis of nutrient deficiencies in rice. Comput. Intell. Neurosci. 2020, 2020, 7307252. [Google Scholar] [CrossRef]
  72. Yu, R.; Huo, L.; Huang, H.; Yuan, Y.; Gao, B.; Liu, Y.; Yu, L.; Li, H.; Yang, L.; Ren, L.; et al. Early detection of pine wilt disease tree candidates using time-series of spectral signatures. Front. Plant Sci. 2022, 13, 1000093. [Google Scholar] [CrossRef] [PubMed]
  73. Lee, J.B.; Kim, E.S.; Lee, S.H. An analysis of spectral pattern for detecting pine wilt disease using ground-based hyperspectral camera. Korean J. Remote Sens. 2014, 30, 665–675. [Google Scholar] [CrossRef] [Green Version]
  74. Qiao, R.; Ghodsi, A.; Wu, H.; Chang, Y.; Wang, C. Simple weakly supervised deep learning pipeline for detecting individual red-attacked trees in VHR remote sensing images. Remote Sens. Lett. 2020, 11, 650–658. [Google Scholar] [CrossRef]
  75. Li, F.; Liu, Z.; Shen, W.; Wang, Y.; Wang, Y.; Ge, C.; Sun, F.; Lan, P. A Remote Sensing and Airborne Edge-Computing Based Detection System for Pine Wilt Disease. IEEE Access 2021, 9, 66346–66360. [Google Scholar] [CrossRef]
  76. Sun, Z.; Ibrayim, M.; Hamdulla, A. Detection of Pine Wilt Nematode from Drone Images Using UAV. Sensors 2022, 22, 4704. [Google Scholar] [CrossRef]
  77. Park, H.G.; Yun, J.P.; Kim, M.Y.; Jeong, S.H. Multichannel Object Detection for Detecting Suspected Trees With Pine Wilt Disease Using Multispectral Drone Imagery. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 2021, 14, 8350–8358. [Google Scholar] [CrossRef]
Remotesensing 15 00444 g001 550
Figure 1. (a) Location of the study area; (b) UAV orthophoto of the study area.
Figure 1. (a) Location of the study area; (b) UAV orthophoto of the study area.
Remotesensing 15 00444 g001
Remotesensing 15 00444 g002 550
Figure 2. (a) DJI Phantom 4 Multispectral Edition; (b) multispectral camera with six images; lenses; (c) an RGB sample along with the reflectance in the different multispectral bands.
Figure 2. (a) DJI Phantom 4 Multispectral Edition; (b) multispectral camera with six images; lenses; (c) an RGB sample along with the reflectance in the different multispectral bands.
Remotesensing 15 00444 g002
Remotesensing 15 00444 g003 550
Figure 3. Temporal changes in the canopy color of infected trees. The text above the date of the picture indicates different stages of infection, and H, GA, E, M, and L represent healthy, green attack, early, middle, and late, respectively.
Figure 3. Temporal changes in the canopy color of infected trees. The text above the date of the picture indicates different stages of infection, and H, GA, E, M, and L represent healthy, green attack, early, middle, and late, respectively.
Remotesensing 15 00444 g003
Remotesensing 15 00444 g004 550
Figure 4. Flowchart of the adopted analysis process.
Figure 4. Flowchart of the adopted analysis process.
Remotesensing 15 00444 g004
Remotesensing 15 00444 g005 550
Figure 5. Temporal variation in tree numbers at different infection stages. The red discontinuous line represents that the number is 20.
Figure 5. Temporal variation in tree numbers at different infection stages. The red discontinuous line represents that the number is 20.
Remotesensing 15 00444 g005
Remotesensing 15 00444 g006 550
Figure 6. Temporal variation in classification accuracy of infected trees (the red discontinuous line from bottom to top represents that WO is 0.7, 0.8, 1.0). The classifications (RF, SVM and LDA) were performed using the basic values and the difference values (indicated by C). RF: random forest; SVM: support vector machine; LDA: linear discriminant analysis. The symbol * indicates significant differences (p < 0.05).
Figure 6. Temporal variation in classification accuracy of infected trees (the red discontinuous line from bottom to top represents that WO is 0.7, 0.8, 1.0). The classifications (RF, SVM and LDA) were performed using the basic values and the difference values (indicated by C). RF: random forest; SVM: support vector machine; LDA: linear discriminant analysis. The symbol * indicates significant differences (p < 0.05).
Remotesensing 15 00444 g006
Remotesensing 15 00444 g007 550
Figure 7. Parameter importance ranking (a) 23 July; (b) 24 August.
Figure 7. Parameter importance ranking (a) 23 July; (b) 24 August.
Remotesensing 15 00444 g007
Remotesensing 15 00444 g008 550
Figure 8. Canopy spectral reflectance at each infection stage.
Figure 8. Canopy spectral reflectance at each infection stage.
Remotesensing 15 00444 g008
Remotesensing 15 00444 g009 550
Figure 9. Weighted accuracy of each infection stage. The red dotted line from bottom to top represents 0.7 and 0.8.
Figure 9. Weighted accuracy of each infection stage. The red dotted line from bottom to top represents 0.7 and 0.8.
Remotesensing 15 00444 g009
Remotesensing 15 00444 g010 550
Figure 10. Parameter importance ranking. The numbers in the picture represent the MDA rank of each parameter.
Figure 10. Parameter importance ranking. The numbers in the picture represent the MDA rank of each parameter.
Remotesensing 15 00444 g010
Remotesensing 15 00444 g011 550
Figure 11. Overview of monitoring period for different PWD infection stage. The patterns indicate data collection dates. Insect patterns indicated peak eclosion time.
Figure 11. Overview of monitoring period for different PWD infection stage. The patterns indicate data collection dates. Insect patterns indicated peak eclosion time.
Remotesensing 15 00444 g011
Table
Table 1. Detailed apparent tree characteristics and classification criteria.
Table 1. Detailed apparent tree characteristics and classification criteria.
PWD Infection StageHealthy StageGreen Attack StageEarly StageMiddle StageLate Stage
UAV (30 m)Remotesensing 15 00444 i001Remotesensing 15 00444 i002Remotesensing 15 00444 i003Remotesensing 15 00444 i004Remotesensing 15 00444 i005
GroundRemotesensing 15 00444 i006Remotesensing 15 00444 i007Remotesensing 15 00444 i008Remotesensing 15 00444 i009Remotesensing 15 00444 i010
Needles Remotesensing 15 00444 i011Remotesensing 15 00444 i012Remotesensing 15 00444 i013Remotesensing 15 00444 i014Remotesensing 15 00444 i015
Table
Table 2. Vegetation indices analyzed in this study.
Table 2. Vegetation indices analyzed in this study.
AbbreviationNameFormulaReference
NDVINormalize difference vegetation indices(ρNIR − ρRed)/(ρNIR + ρRed)[34]
NDRENormalize difference red-edge indices(ρNIR − ρRed Edge)/(ρNIR + ρRed Edge)[35]
GLIGreen leaf index2 × (ρGreen − ρRed − ρBlue)/2 × (ρGreen + ρRed − ρBlue)[36]
CIGChlorophyll index green(ρNIR/ρGreen) − 1[37]
CVIChlorophyll vegetation indexρNIR(ρRed/ρGreen2)[38]
NGRDINormalize difference Green/red(ρGreen − ρRed)/(ρGreen + ρRed)[39]
PBIPlant biochemical index(ρNIR)/(ρGreen)[40]
GNDVIGreen normalized difference vegetation index(ρNIR − ρGreen)/(ρNIR + ρGreen)[41]
LCILeaf chlorophyll index(ρNIR − ρRed Edge)/(ρNIR + ρRed)[42]
RVIRatio vegetation indexρNIR/ρRed[43]
EVIEnhanced vegetation index 2.5(ρNIR − ρRed)/(ρNIR + 6ρRed − 7.5ρBlue + 1)[44]
DVIDifference vegetation index ρNIR − ρRed[45]
RDVIRe-normalized difference vegetation indexSQRT (NDVI × DVI)[46]
TVITriangular vegetation index 60(ρNIR − ρGreen) − 100 (ρRed − ρGreen)[47]
VARIVegetation atmospherically resistant index(ρGreen − ρRed)/(ρGreen+ρRed − ρBlue)[48]
NWINormalized wilt indexNWI = −NDGI × (NDVI + NDGI)
NDGI = (ρRed − ρGreen)/(ρRed + ρGreen)		[49]
PSRIPlant Senescence Reflectance Index(ρRed − ρBlue)/ρRed Edge[50]
BRBlue ratio(ρRed/ρBlue) × (ρGreen/ρBlue) × (ρRed Edge/ρBlue) × (ρNIR/ρBlue)[51]
Note: ρi is the reflectance at wavelength i and SQRT is square root.
Table
Table 3. The difference comparison results of the accuracy of the three algorithms under each date.
Table 3. The difference comparison results of the accuracy of the three algorithms under each date.
DateTukey’s Multiple Comparisons TestMean Diff.95.00% CI of Diff.Adjusted p Value
23 JulyRF vs. SVM−0.006733−0.05320 to 0.039730.9374
RF vs. LDA0.04377−0.002696 to 0.090240.0695
SVM vs. LDA0.050500.004037 to 0.096970.0295
10 AugustRF vs. SVM−0.007937−0.05440 to 0.038530.9141
RF vs. LDA0.03968−0.006784 to 0.086150.1107
SVM vs. LDA0.047620.001153 to 0.094090.0432
17 AugustRF vs. SVM0.01961−0.02686 to 0.066070.5792
RF vs. LDA−0.01961−0.06607 to 0.026860.5792
SVM vs. LDA−0.03922−0.08568 to 0.0072500.1164
24 AugustRF vs. SVM0.03292−0.01354 to 0.079390.2177
RF vs. LDA0.063790.01732 to 0.11030.0040
SVM vs. LDA0.03086−0.01560 to 0.077330.2612
2 SeptemberRF vs. SVM0.002137−0.04433 to 0.048600.9935
RF vs. LDA−0.04274−0.08920 to 0.0037310.0785
SVM vs. LDA−0.04487−0.09134 to 0.0015940.0609
22 SeptemberRF vs. SVM0.02469−0.02177 to 0.071160.4218
RF vs. LDA0.067020.02055 to 0.11350.0023
SVM vs. LDA0.04233−0.004138 to 0.088790.0823
12 OctoberRF vs. SVM−0.004831−0.05130 to 0.041640.9672
RF vs. LDA−0.02415−0.07062 to 0.022310.4376
SVM vs. LDA−0.01932−0.06579 to 0.027140.5884
Table
Table 4. Confusion matrix for several key days.
Table 4. Confusion matrix for several key days.
Date23 July10 August24 August
Infection stagesUAWOKAPPAUAWOKAPPAUAWOKAPPA
Healthy0.7860.6360.3880.9230.7140.6111.0000.7590.671
Green attack0.6430.2500
Early0.2500.6150.625
MiddleNA0.7000.824
Late01.0000.625
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Wu, D.; Yu, L.; Yu, R.; Zhou, Q.; Li, J.; Zhang, X.; Ren, L.; Luo, Y. Detection of the Monitoring Window for Pine Wilt Disease Using Multi-Temporal UAV-Based Multispectral Imagery and Machine Learning Algorithms. Remote Sens. 2023, 15, 444. https://doi.org/10.3390/rs15020444

AMA Style

Wu D, Yu L, Yu R, Zhou Q, Li J, Zhang X, Ren L, Luo Y. Detection of the Monitoring Window for Pine Wilt Disease Using Multi-Temporal UAV-Based Multispectral Imagery and Machine Learning Algorithms. Remote Sensing. 2023; 15(2):444. https://doi.org/10.3390/rs15020444

Chicago/Turabian Style

Wu, Dewei, Linfeng Yu, Run Yu, Quan Zhou, Jiaxing Li, Xudong Zhang, Lili Ren, and Youqing Luo. 2023. "Detection of the Monitoring Window for Pine Wilt Disease Using Multi-Temporal UAV-Based Multispectral Imagery and Machine Learning Algorithms" Remote Sensing 15, no. 2: 444. https://doi.org/10.3390/rs15020444

APA Style

Wu, D., Yu, L., Yu, R., Zhou, Q., Li, J., Zhang, X., Ren, L., & Luo, Y. (2023). Detection of the Monitoring Window for Pine Wilt Disease Using Multi-Temporal UAV-Based Multispectral Imagery and Machine Learning Algorithms. Remote Sensing, 15(2), 444. https://doi.org/10.3390/rs15020444

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop