Study area
Sanhe National Forest Farm (Fig.1), with a total area of 247.6 hectares, located in Qingshui town of Jiuquan city, Gansu Province, China, was selected as the experimental plot. The region was the first to be invaded by ALB in Jiuquan city and, therefore, it contains abundant samples demonstrating different damage stages. Because it is an intensively managed forest farm, the tree species is relatively single, which is conducive to our research. The local altitude is 1480 m, with the following geographical coordinates: east longitude of 98°20′-99°18′ and north latitude of 39°10′-39°59′. The annual average precipitation is 80 mm, but the annual average evaporation reaches 2000 mm. The annual average temperature is 4°C–6°C, and the monthly average temperature is -15℃–29℃.
Ground truth data collection
Ground truth data were collected in September 2020. The hazardous area of ALB infestation was determined through extensive inspection. After selecting the most suitable area for the study, 229 sample trees with similar breast diameters and different damage degrees were randomly selected. Dead trees were not considered for the following reasons: (1) Dead trees do not contain ALBs and identifying them for a study on infection prevention is meaningless. (2) The canopy of a dead tree is small and typically short, which presents challenges in distinguish them using WV-2 images or dead trees are completely hidden by taller trees, hindering visualization from a higher point. Forest pest management often requires damage quantification. To determine the damage stages of P. gansuensis, we clicked photos in four directions and measured the diameter at breast height (DBH), global positioning system (GPS) location, exit hole level, dead twig level, damaged ratio, and leaf area index (LAI) of each inspected tree to determine the damage stages at the individual level. We also obtained orthographic images of each inspected P. gansuensis by using unmanned aerial vehicles (UAVs). Fig. 2 presents examples of different ALB-induced damage stages of P. gansuensis.
Leaf area index
LAI characterizes the leaf density and canopy structure to reflect the ability of photosynthesis, respiration, transpiration, and other biophysical processes of the vegetation (LIU Yang et al. 2013). The average LAI value is the mean of individual tree measurements from four directions by using LAISmart (Francesca Orlando et al. 2015).
Exit hole level
Exit hole is an important symptom of ALB infestation. Exit holes from larvae are the most obvious and direct evidence of ALB infestation (Y. Fragnière et al. 2017). We examined the main holes of each tree, and telescopically counted the number of exit holes as an estimate of ALB activity. Exit holes are divided into level 1-5 from less to more according to quantity.
Dead twig level
Dead twig level is based on the visual assessment of the overall condition of individual trees. Level values range from 1 to 5, according to a technical scheme of Chinese forest pest survey (Chinese National Forestry and Grassland Administration, 2014). The level of dead branches refers to the ratio of dry branches to the total number of branches, with 1 indicating a healthy branch (no major twig mortality), 2 indicating a slight decline in the tree’s health (1%–30% crown damage), 3 indicating a moderate decline in the tree’s health (30%–60% crown damage),4 indicating severe decline in the tree’s health (>60% crown damage), and 5 indicating a dead twig.
Damage shoot ratio
Shoot damage index is a widely used and an important indicator of the extent of injury to a tree (Qinan Lin et al. 2019). It refers to the ratio of large shoots with symptoms of damage to the total number of large shoots. Shoot damage data are also obtained through visual assessment by forestry experts according to the technical scheme of Chinese forest pest survey (Chinese National Forestry and Grassland Administration, 2014). The index value ranges from 0% (indicating a completely healthy tree) to 100% (indicating a dead tree).
Image acquisition using UAV
Images were acquired using Yu Mavic2 professional(DJI , China). The orthographic images of single inspected trees were clicked from a height of 50–100 m. The 50-m-high orthographic images were printed to map the inspected individual tree for determining the exact location of each inspected tree. UAV images were acquired to determine the canopy color of individual trees and to serve as a supplement sub-meter GPS information for analyzing the accurate position of the inspected tree in the forest.
Decline rating summary
To determine the damage stages of P. gansuensis and the reference basis for remote-sensing images, we standardized and integrated the canopy color and aforementioned individual physiological indices according to the experience of practical investigation and previous research on coniferous trees and explored their relevance. Instead of using a numerical value to indicate the damage stage of a broad-leaved tree according to previous studies, we used canopy color, which has a strong correlation with a tree’s health and can be assessed directly from remote-sensing data, as a criterion.
The following verification was performed to determine the canopy color: First, the corresponding canopy color of inspected trees was determined from the UAV image, followed by comparison of the average and standard deviation of the tree’s physiological factors corresponding to tree canopy colors. Second, the canopy color was used as an independent variable to perform a one-way ANOVA involving the tree’s physiological factors. A p value of <0.05 was considered statistically significant. Statistical analyses were performed using R (R Development Core Team, 2018).
Satellite image acquisition
Considering the accuracy requirements for distinguishing single trees and actually available satellite data sources, we obtained the corrected WV-2 commercial 8-band very high resolution (VHR) satellite images on September 27, 2020. The WV-2 images of the study area were ordered as multispectral and panchromatic, which contains 0% cloud cover and spatial resolutions of 2 m (multispectral) and 0.5 m (panchromatic). Main sensor specifications are described in Table 1.
The satellite scene was coded in units of numbers (DN). Calibration and atmospheric correction models (FLAASH in ENVI 5.3) were applied to the multispectral image to convert the digital (DN) value to the sensor radiation and reflectance values. We used the Gram-Schmidt Pan-sharpening technology to pan-sharpen multispectral images through the full color band. Finally, a 5-m digital terrain model was used to orthorectify the 0.5-m fully sharpened multispectral image.
Table.1 Technical specifications of the WV-2 imagery
Senor characteristic
|
Spatial Resolution
|
0.5m PAN and 2m MS
|
Spectral Resolution [nm]
|
Coastal: 400–450
|
Blue: 450–510
|
Green: 510–580
|
Yellow: 585–625
|
Red: 630-690
|
Red Edge: 705–745
|
NIR1: 770–895
|
NIR2: 860–1040
|
Data acquisition
|
Jiuquan city, 27 Sep 2020
|
Tree crown segmentation
Methodology schema with all steps applied in classification of P. gansuensis damage stages is showed in Fig.3. We used pan-sharpened WV-2 images and compared their accuracy to detect damaged branches of the affected trees at different levels according to the object-based method. We used eCognition Developer 9.0 (Trimble Geospatial, USA) to subdivide the pan-sharpened WV-2 images into image objects through multiresolution segmentation. To keep the canopy consistent with the line segment polygons, we iteratively used multiple subdivision levels of detail to adapt to the shape and tightness parameters. The multiresolution classification method was used to categorize the images. The segmentation steps were as follows: initial segmentation of a single tree: scale parameter, 1; shape, 0.7; compactness, 0.9; de-shading, 5.6e-34 < intensity < 7.6e-34; forest area classification: NDVI > 0.26; single tree segmentation: scale parameter, 4; shape, 0.6; compactness, 0.6.
Reference data modeling and predicting
In order to ensure that the sample size of different degrees of damage is roughly the same, except for the sample trees in the field survey, we also visually inspected the canopy color according to the UAV image. We manually extracted data on tree crowns to construct a spectral diagnostic model, followed by applying the model to the entire satellite image to distinguish the damage stages of individual trees. A total of 139, 139, and 121 trees were in the green stage, yellow stage, and gray stage, respectively. Further, the reflectance of each canopy of the eight bands was extracted to calculate various vegetation indices (VIs). Finally, the reflectance of 8 bands and 17 types of VIs were extracted to construct the database, as well as the training and validation sets at a 7:3 ratio. The extracted VIs are presented in Table 2. On the basis of WV-2 data, the classification model demonstrated the highest classification accuracy, and the Kappa index was adopted to generate a prediction map at the individual tree level.
Table.2 Remote sensing vegetation indices tested in this study and adapted to the WV-2
Abbreviation
|
Name
|
Formula
|
ref
|
NDVI
|
Normalized Difference Vegetation Index
|
(NIR1 − R)/(NIR1 + R)
|
Zhang, X,2006
|
NDVI3,5
|
Green–red ratio
|
(G − R)/(G + R)
|
Gitelson,A.A.1996
|
NDVI8,4
|
NIR-yellow ratio
|
(NIR2 − Y)/(NIR2 + Y)
|
Gwata, B.2012
|
NIRRY
|
NIR-Red-yellow ratio
|
(NIR1)/(R + Y)
|
Gwata, B.2012
|
DD
|
Difference Difference Vegetation Index
|
(2 ×NIR1−R)−(G−B)
|
Le Maire, G.;2004
|
NORM NIR
|
Normalized NIR
|
NIR1/(NIR1 + R + G)
|
ENVI, 2013
|
PSRI
|
Plant Senescence Reflectance Index
|
(R − B)/RE
|
Sims, D.A.,2002
|
RVI
|
Ratio vegetation index
|
NIR/RED
|
Hildebrandt,G.1996
|
GR
|
Green–red ratio
|
G/R
|
Lars T. Waser,2014
|
BR
|
Blue ratio
|
(R/B) ×(G/B) ×(RE/B) ×(NIR/B)
|
Lars T. Waser,2014
|
RR
|
Red ratio
|
(NIR1/R) ×(G/R) ×(NIR1/RE)
|
Lars T. Waser,2014
|
REY
|
RedEdge yellow ratio
|
(RE − Y)/(RE + Y)
|
Gwata, B.2012
|
VIRE
|
Vegetation Index based on RedEdge
|
NIR1/RE
|
Chávez, R,2011
|
RGI
|
Red–green index
|
RED/GREEN
|
Miura, T,2008
|
EVI1
|
Enhanced Vegetation Index 1
|
2.4*(NIR1−RED)/(NIR1+RED+1)
|
F.G.S.Bezerra,2020
|
EVI2
|
Enhanced Vegetation Index 2
|
2.4*(NIR2−RED)/(NIR2+RED+1)
|
Jiang, Z.2008
|
GI2
|
Greenness Index 2
|
(B×−0.2848+G×−0.2434+R ×−0.5436 + NIR1 ×0.7243 + NIR2 ×0.0840) ×5
|
Gwata, B.2012
|
Variable selection and data modeling
In total, data of 399 individual trees at three damage stages were used to construct a classification model among WV-2 images. First, to select the best explanatory variable among the 8 original bands and 18 VIs, a variance inflation factor (VIF) analysis to check for multicollinearity was performed. A VIF value of ≥10 indicated the need to eliminate serious collinearity. Second, a stepwise regression analysis was performed. The analysis was based on the Akaike information criterion (AIC) information statistics as the criterion by selecting the smallest AIC information statistics to achieve a purpose of deleting or adding variables. Lastly, we applied three machine learning algorithms to classify P. gansuensis into three damage stages: Random Forest(RF), Support Vector Machine(SVM)and Classification And Regression Tree(CART). RF and SVM have been widely used in single wood damage classification and have shown good performance, whereas CART can be easily implemented and explained by certain rules (Jing, W et al. 2015; Kaszta, Z et al. 2016).
RF is an improved algorithm compared to traditional decision trees that generates numerous decision trees. Among the classification results of all the constructed decision trees, new data are classified based on the majority of votes (Breiman, L et al. 2001). The SVM algorithm helps find the best hyperplane as the decision function in the high-dimensional space and classify the input vector into different classes (Cortes, C et al. 1995). CART is a binary recursive partitioning algorithm based on tree nodes generated by training data (Everitt, B.S et al. 2005). Finally, the overall accuracy (OA), producer’s accuracy, user’s accuracy (UA), and the Kappa coefficient generated by the confusion matrix were used to evaluate the accuracy of identification of the damaged stage of a single tree. Kappa values of <0.4, 0.4-0.8, and >0.80 indicate poor agreement, moderate agreement, and strong agreement, respectively (Arjan J.H. Meddens et al. 2011).