1. Introduction
Eucalyptus Longhorned Borers (ELB),
Phoracantha semipunctata (Fabricius), and
P. recurva Newman (Coleoptera:
Cerambycidae), are among the most destructive eucalypt pests in regions with Mediterranean climate [
1,
2].
Eucalyptus globulus is one of the most planted eucalypt species in these regions, and it is known to have low resistance to ELB [
2,
3].
ELB activity generally starts in late spring when adults emerge and start laying eggs. Upon hatching, ELB larvae bore galleries along the phloem and cambium of trees, eventually preventing sap from flowing, which leads to rapid tree death during summer and fall [
1,
4,
5,
6]. The ability of larvae to successfully colonize the host plant depends on low bark moisture content, leaving water-stressed trees particularly susceptible to attack by ELB [
1,
5,
7].
Under the Mediterranean climate, with low rainfall and extended dry summers periods, the attack by ELB often result in significant tree mortality, despite long-lasting efforts to select for more resistant
E. globulus genotypes [
8]. With droughts expected to increase due to climate change, ELB outbreaks will likely become more frequent and more severe [
9].
ELB control methods include biological control, selection of more resistant eucalypts, and various cultural practices aimed at increasing tree adaptation and resilience, but the most effective curative measure is felling all attacked trees and removing them from the stands [
6,
10]. Damage caused by ELB is usually not detected using traditional surveillance techniques until significant mortality has occurred [
11]. Early detection is recognized as the first step to reduce the impact of ELB [
1], hence new approaches to monitoring, particularly through remote sensing, can provide an invaluable tool for forest managers in terms of planning and executing control actions.
Traditional survey techniques are restricted by small area coverage and subjectivity [
12] but combined with remote-sensing technology can lead to expanded spatial coverage, minimize the response time, and reduce the costs of monitoring forested areas [
13]. Following Wulder et al. [
14], the appropriate sensor and resolution should be adopted according to the spatial scale which better adjusts to the situation. To date, several studies have demonstrated successful pest and diseases detection and monitoring in forests using different types of sensor and platform [
15,
16,
17]. For instance, in terms of spaceborne optical sensors for large areas, Meddens et al. [
18] detected multiple levels of coniferous tree mortality using multi-date and single-date Landsat imagery. Mortality in ash trees was assessed by Waser et al. [
19] using multispectral WorldView-2 imagery. In the eucalypt forest context, to predict bronze bug damage Oumar and Mutanga [
20] tested WorldView-2 imagery. The airborne optical sensor Compact Air-Borne Spectrographic Imager 2 (CASI-2) has been tested by Stone et al. [
21] who conducted a study to assess damage caused by herbivorous insects in Australian eucalypt forests and pine plantations.
In recent years the use of unmanned aerial vehicle (UAV) platforms has become widely employed for pests and disease detection and monitoring [
12,
22,
23,
24,
25,
26,
27,
28]. The main attributes are very high resolution, suitability for multitemporal analysis, lower operational costs compared with airplanes and satellites, independence of cloud cover, and the ability to operate in specific phenological phases of plants or pest/disease outbreaks [
12]. In addition, a large number of passive and active sensors can be assembled, such as RGB (red, green, blue), multispectral and hyperspectral cameras, LiDAR (laser imaging detection and ranging), and RADAR (radio detection and ranging) [
29,
30]. On the other hand, as stressed by Pádua et al. [
29], the disadvantages include small area-coverage when compared with satellites, sensitivity to bad weather, increasingly strict regulations that may restrict operations, and the high volume of data generated.
The imagery classification approaches used in the more recently published UAV studies to detect pests and diseases in forests are diverse. Lehmann et al. [
22] used a modified normalized difference vegetation index (NDVI) to discriminate between five classes of infestation by the oak splendor beetle through OBIA (objected-based imagery analysis) classification with an overall Kappa index of agreement of 0.81–0.77. Näsi et al. [
23] investigated bark beetle damage at the tree level in South Finland using a combination of RGB, near infrared (NIR), red-edge band and NDVI. Object-based K-NN (K-nearest neighbor) classification was applied, and overall accuracy was 75%. In New Zealand, Dash et al. [
12] used NDVI and red-edge NDVI to study the discoloration classes of
Pinus radiata through the OBIA classification approach with a random forest (RF) algorithm. In Catalonia (Spain), Otsu et al. [
26] detected defoliation of pine trees affected by pine processionary and distinguished pine species at a pixel level using four spectral indices and NIR band. The authors estimated the threshold values using histogram analysis. For comparisons, another classification used was the OBIA with a random forest algorithm and an overall accuracy of 93%. Finally, Iordache et al. [
28] studied pine wild disease in Portugal using OBIA and machine-learning random forest algorithm for multispectral and hyperspectral imagery. The overall accuracy of both classifications was 95% and 91%, respectively. To date, no published work has been developed with UAV imagery for ELB detection.
This work arises from the necessity to find monitoring tools that could support pest management decisions regarding ELB attacks in eucalypt stands. In the past, this has been hampered since the identification of dead trees in the field is bothcostly and time-consuming. Furthermore, it is critical that ELB attacks can be identified at early stages of infestation. Current surveying methods only detect problems when large number of trees are infested.
The main objectives of this experimental work were: (1) to detect ELB attacks through UAV imagery by using selected spectral indices and Otsu thresholding; (2) to map trees crown status using large-scale mean-shift (LSMS) segmentation, as well as the machine-learning classification approach with a RF algorithm; and (3) to map tree density using the hexagonal tessellations technique to support phytosanitary interventions.
4. Discussion
This study explored the capability of multispectral images captured with a parrot sequoia camera to detect the mortality caused by ELB on E. globulus plantations. Before establishing a flight plan, the phenology of the trees and the life cycle of the insect must be known, in order to extract as much information as possible. The spatial resolution of the images obtained was 17 cm, which allowed the reduction of spatial scale to the individual tree level.
NIR bands had the greatest discriminating power of the all spectra analyzed and provided more information about different tree canopies. As stressed by Otsu et al. [
26], NIR band was more sensitive to defoliation than the red-edge band, as observed in
Figure 7b. Regarding the remaining bands, the reflectance of the green band was generally higher than that of red band because of the high absorption related to chlorophyll content [
84]. Spectral indices, whose calculation includes bands such as NDVI, also presented highly discriminating behavior. This VI is particularly used to assess defoliation, and damage at high spatial resolutions using multispectral and RGB cameras mounted on UAV [
12,
22,
23,
26,
85,
86].
Histogram analysis allowed the discrimination between healthy and dead trees. Although the detection of the histogram valley in spectral indices was difficult to determine, this method may become a valid alternative to split the two canopy types. NDVI was the most accurate index at 98.2% and Kappa value of 0.96. The second most accurate was the GNDVI index, which had an overall accuracy of 97.4% and Kappa value of 0.95. NDRE was the less accurate index at 84.4% overall accuracy and Kappa value of 0.73. Similar overall accuracy values were obtained by Otsu et al. [
26] to detect defoliation of pine needles by pine processionary in four different locations. However, in our study, the NIR band was not used to remove shadows as Miura and Midokawa et al. [
87] applied. To minimize the shadow effect, the flight was undertaken at solar noon, so treetops could be easily captured, with shadows present mostly along the plantation lines.
The combination of segmentation and the maximum filtering location allowed the extraction of the tops of trees and carrying out their classification. Applying this method, shadows were removed from other vegetation and bare soil, which has a reflectance very close to that of dead trees. The RF learning machine obtained 98.5% global accuracy and the Kappa value was 0.94. This precision is explained by the great distance between reflectance values and the differences between the two classes. Iordache et al. [
28] applied RF classifier on the classification of
Pinus pinaster canopy types (infected, suspicious, and healthy) affected by pine wild and obtained an overall accuracy of 95%. Pourazar et al. [
27], obtain as overall accuracy 95.58% using five spectral bands and five indices to detect dead and diseased trees.
The forest density maps produced through hexagon tessellations aimed to group the position of classified trees. The ease of reading allows the identification of the most critical areas with tree mortality and to extract important metrics for forest management such as the total number of trees, the standard deviation, and other landscape metrics. Barreto et al. [
88] set a hexagonal grid to represent classes of natural Cerrado vegetation in the the Northeast of Brazil, in an area of about 25,590 km
2, to study the remaining habitats through quantitative indices. More recently, Amaral et al. [
89], used a hexagonal grid subdivided into 1000 ha units to study the
restinga (a type of coastal vegetation) in Northern Brazil.
The current strategy to control ELB relies mostly on the identification of dead trees and their removal from plantations before a new generation of adults emerges in late spring. Traditional field surveys are extremely labor-intensive, as they require visual assessment of large eucalypt plantations. As a result, outbreaks are often only detected when large numbers of trees are already dead and ELB populations are high. By allowing the identification of individual infested trees in a much more efficient way, widespread application of the methods described in this study will allow forest managers to detect ELB attacks at an early stage, thus reducing the cost and efficiency of sanitary fellings.
Future work will focus on adding additional classes to the survey in order to improve the discrimination of tree health status. The integration of a UAS-derived canopy model at a 3D tree model could be performed to automatically outline individual tree crowns. Another improvement is the ambition to implement periodical flights at different attack stages to provide a multitemporal analysis. Further research might focus on other remote-sensing platforms at different scales and considering different bioclimatic and geographical settings.