Early Monitoring of Health Status of Plantation-Grown Eucalyptus pellita at Large Spatial Scale via Visible Spectrum Imaging of Canopy Foliage Using Unmanned Aerial Vehicles

: Eucalyptus is a diverse genus from which several species are often deployed for commercial industrial tree plantation due to their desirable wood properties for utilization in both solid wood and ﬁber products, as well as their growth and productivity in many environments. In this study, a method for monitoring the health status of a 22.78 ha Eucalyptus pellita plantation stand was developed using the red-green-blue channels captured using an unmanned aerial vehicle. The ortho-image was generated, and visual atmospheric resistance index (VARI) indices were developed. Herein, four classiﬁcation levels of pest and disease were generated using the VARI-green algorithm. The range of normalized VARI-green indices was between − 2.0 and 2.0. The results identiﬁed seven dead trees (VARI-green index − 2 to 0), ﬁve trees that were severely infected (VARI-green index 0 to 0.05), 967 trees that were mildly infected (VARI-green index 0.06 to 0.16), and 10,090 trees that were considered healthy (VARI-green index 0.17 to 2.00). The VARI-green indices were veriﬁed by manual ground-truthing and by comparison with normalized difference vegetation index which showed a mean correlation of 0.73. This study has shown practical application of aerial survey of a large-scale operational area of industrial tree plantation via low-cost UAV and RGB camera, to analyze VARI-green images in the detection of pest and disease.


Introduction
Plantations of genus Eucalyptus L'Hér now amount to more than 20 million hectares globally [1] and are the second largest global forest plantation species behind Pinus L. [2]. Their capability for fast growth, ability to grow in a range of site conditions, ease of propagation, and their desirable wood qualities have led to widespread establishment of large Eucalyptus plantations in many countries of Southeast Asia which has over 2.5 million planted ha [3]. The wood of Eucalyptus is suitable for use in a variety of products, such as sawn timber, pulp, paper, civil construction, furniture and energy production purposes, and veneer or plywood production. In the Malaysian state of Sabah on Borneo Island, Eucalyptus species and hybrids have been shown to have high productivity and have a variety of end use potentials [4][5][6][7]. As countries increasingly restrict the access to natural forests for log supply, there are estimates that 50% of the world's hardwood harvest could

Study Area and Tree Health Data
This study was undertaken in the industrial tree plantation of Brumas estate at Sabah Softwoods Berhad (Tawau, Sabah, located at latitude 4 • 35 36 N and longitude 117 • 45 31 E between 200 and 600 m elevation above sea level in the southern regionof the Malaysian state of Sabah on the island of Borneo (Figure 1). The total planted area at SSB is approximately 18,000 hectares, which is predominantly E. pellita (60%) and Falcataria moluccana Barneby & Grimes (formerly Albizia falcataria L. Fosberg) as a secondary species (25%), along with Eucalyptus hybrids (5%) and conservation plantings (10%). The soil is dominated by Tanjung Lipat soil type with clay texture between 25% to 35% and Kumansi type with >40% clay content [40]. Block 42H, a commercial planting of E. pellita and selected for this study, was planted in May 2018, with an area of 22.78 ha at a spacing of 3 m × 3 m (1111 stems per hectare, sph). A younger commercial planting (six-month-old E.pellita) was also assessed in Block 42G. In this block, a total of 4399 trees were planted in an area of 7.56 ha. visual inspection, and (3) develop a range of indices of VARI-green as benchmark for detection and monitoring of health status.

Study Area and Tree Health Data
This study was undertaken in the industrial tree plantation of Brumas estate at Sabah Softwoods Berhad (Tawau, Sabah, located at latitude 4°35′36″ N and longitude 117°45′31″ E between 200 and 600 m elevation above sea level in the southern regionof the Malaysian state of Sabah on the island of Borneo (Figure 1). The total planted area at SSB is approximately 18,000 hectares, which is predominantly E. pellita (60%) and Falcataria moluccana Barneby & Grimes (formerly Albizia falcataria L. Fosberg) as a secondary species (25%), along with Eucalyptus hybrids (5%) and conservation plantings (10%). The soil is dominated by Tanjung Lipat soil type with clay texture between 25% to 35% and Kumansi type with >40% clay content [40]. Block 42H, a commercial planting of E. pellita and selected for this study, was planted in May 2018, with an area of 22.78 ha at a spacing of 3 m × 3 m (1111 stems per hectare, sph). A younger commercial planting (six-month-old E.pellita) was also assessed in Block 42G. In this block, a total of 4399 trees were planted in an area of 7.56 ha.

Early Inventory Measurement
The early inventory measurement (EIM) was completed five months after planting, while the ground-truth verification was completed after completion of the flight used for the VARI-green analysis.
In the EIM, visual inspection of E. pellita health was conducted at all study sites to establish ground reference dataset. Tree health inspection was undertaken by an experienced assessor who was trained and actively engaged in pest and disease survey in Sabah Softwood plantations.
The block 42H was chosen because of the young age when assessed using an unmanned aerial vehicle in November 2019 (1 year 7 months [19 months]), and the early inventory measurement (EIM) on the block which showed a variety of tree health status from virulent to dead.

Early Inventory Measurement
The early inventory measurement (EIM) was completed five months after planting, while the ground-truth verification was completed after completion of the flight used for the VARI-green analysis.
In the EIM, visual inspection of E. pellita health was conducted at all study sites to establish ground reference dataset. Tree health inspection was undertaken by an experienced assessor who was trained and actively engaged in pest and disease survey in Sabah Softwood plantations.
The block 42H was chosen because of the young age when assessed using an unmanned aerial vehicle in November 2019 (1 year 7 months [19 months]), and the early inventory measurement (EIM) on the block which showed a variety of tree health status from virulent to dead.

Unmanned Aerial Vehicle Image Acquisition
High-resolution aerial images were acquired using an unmanned aerial vehicle (UAV) namely a Phantom 4 Pro, (Shenzen DJI Sciences and Technologies (DJI), Shenzhen, China, Figure 2) equipped with a readily available consumer market RGB camera (Go-Pro 5, San Mateo, CA, USA) with a payload weight of 500 g. This configuration of UAV was used to obtain high spatial resolution imagery in the visible spectrum. The aerial survey work was undertaken on 27 November 2019, with an altitude above canopy of 129 m and overlapping image swathes of 60% (forward) and 30% (side). Map Pilot for DJI software (Drones Made Easy, San Diego, CA, USA) was used to set the parameters (Table 1) for flightpath planning and control the movement and speed of the UAV. The constant ground sample distance (GSD) mode was set during the flight. A total of 178 images were obtained. The ground truth (GT) covered 24 plots (radius of 11.5 m) randomly scattered in the block which represents 4.22% of the total area (22.87 ha).

Unmanned Aerial Vehicle Image Acquisition
High-resolution aerial images were acquired using an unmanned aerial vehicle (UAV) namely a Phantom 4 Pro, (Shenzen DJI Sciences and Technologies (DJI), Shenzhen, China, Figure 2) equipped with a readily available consumer market RGB camera (Go-Pro 5, San Mateo, CA, USA) with a payload weight of 500 g. This configuration of UAV was used to obtain high spatial resolution imagery in the visible spectrum. The aerial survey work was undertaken on 27 November 2019, with an altitude above canopy of 129 m and overlapping image swathes of 60% (forward) and 30% (side). Map Pilot for DJI software (Drones Made Easy, San Diego, CA, USA) was used to set the parameters (Table 1) for flightpath planning and control the movement and speed of the UAV. The constant ground sample distance (GSD) mode was set during the flight. A total of 178 images were obtained. The ground truth (GT) covered 24 plots (radius of 11.5 m) randomly scattered in the block which represents 4.22% of the total area (22.87 ha).  Ground truth assessments were conducted to classify infestation levels on E. pellita caused by pests and diseases. Tree health status was based on the following conditions: healthy (no sign of infestation on foliage and stem), moderate (distortion in foliage and shoots dieback), severe (splitting of stem with distortion and/or discoloration in foliage), and dead (trees no longer functional with discoloration of vascular tissues and wilting of foliage). These severity levels were used to demonstrate the potential of RGB remote images to distinguish health status of individual trees.  Ground truth assessments were conducted to classify infestation levels on E. pellita caused by pests and diseases. Tree health status was based on the following conditions: healthy (no sign of infestation on foliage and stem), moderate (distortion in foliage and shoots dieback), severe (splitting of stem with distortion and/or discoloration in foliage), and dead (trees no longer functional with discoloration of vascular tissues and wilting of foliage). These severity levels were used to demonstrate the potential of RGB remote images to distinguish health status of individual trees.

Image Processing
Images were processing using Agisoft Metashape software (Agisoft, St Petersburg, Russia) to generate ortho-images using parameters shown in Table 2, which is this similar to the parameters used by Del-Campo-Sanchez et al. [18]. Three key steps were used to process the UAV images: initial processing, point cloud densification, and ortho-image generation. For initial processing, an automatic aerial triangulation (AAT) algorithm was used to refine the exterior orientation for all images (N = 178) in order to compute direct georeferencing for each image. Then, bundle block adjustment (BBA) was automatically performed to utilize the measured distance between points in order to generate an adjusted relative three-dimensional (3D) coordinate system. For point cloud densification (Figure 3), the X, Y, Z position and the color information were computed based on the automatic tie points (ATP) in the initial processing stage. In ortho-image generation, digital surface models (DSM) with resolution 0.135 m were generated and used as reference for orthomosaic image creation. Agisoft software was used to import a third-party surface model for ortho-image generation, light detection and ranging (LIDAR), or interferometry synthetic aperture radar (IFSAR). The geomatic information obtained by GSD photogrammetric techniques is in this study a minimum of 3.5 cm pixel −1 . The tie points for the cloud characteristics are summarized in Table 3. The setting during the flight plan met the geomatic proposed desirable quality of final ortho-mosaic product. The ortho-images covered 27.22 hectares and provided information at a resolution of 3.5 cm GSD. The volume of data obtained from this setting were summarized in Table 4.  The geomatic information obtained by GSD photogrammetric techniques is in this study a minimum of 3.5 cm pixel −1 . The tie points for the cloud characteristics are summarized in Table 3. The setting during the flight plan met the geomatic proposed desirable quality of final ortho-mosaic product. The ortho-images covered 27.22 hectares and provided information at a resolution of 3.5 cm GSD. The volume of data obtained from this setting were summarized in Table 4.

Image Analysis
Image analysis was performed using ArcGIS version 10.1 (Esri, Redlands, CA, USA). The UAV images were converted into vector shapefiles (.shp), after which, point location of the trees was plotted and buffered at a 1.0 m radius using the editor and multiple ring buffer tools. Forest blocks, roads, and stream features in the study area were digitized to represent the complete land cover map of Block 42H. The digitized tree was buffered to 1.0 m radius as an estimation of the Eucalyptus trees crown diameter affected by disease or mortality. The estimation is based on the maximum healthy tree crown diameter in Block 42H. The buffering layer were then overlaid and clipped on the VARI-green image using Extract tools. The VARI-green index is a vegetation index produced entirely in the visible region of the spectrum using the three visible channels of red (R, 564-580 nm), green (G, 534-545 nm), and blue (B, 420-440 nm) using the algorithm of [32] ((G − R)/(G + R − B)). The index generates a greenness level, from which tree health can be determined. The higher the index level, the healthy the trees.
The processed image bands were separated into single bands of R, G, and B using "make a raster layer" in management tools. Then, the single band inserted in raster calculator using syntax dialog. The VARI-green index was computed using Spatial Analyst of Raster calculator tool then, the value was normalized with an absolute value 10 (ABX) and log was set to Log 10 . After the soil threshold was determined, the VARI-green image was clipped using Data Analysis tools to isolate the individual trees from soil feature. Total digital value in the VARI-green image was normalized using ABX and LOG formula. Log 10 was used to raise the digital number. The range between −2 and 2 represents the damage severity levels for infected trees (light green) to healthy trees (deep green).

Zonal Statistics
Once separation between trees and soil on the VARI-green raster image is completed, the calculations of digital number were made to identify the healthy and unhealthy trees. The Zonal Statistic tools in ArcGIS 10.1 were used to calculate the digital number on the VARI-green image. With zonal statistic tools, all digital numbers on the VARI-green were calculated based on each zone dataset. A single output value is computed for every zone in the input dataset. The statistic input was recorded using Excel files for every tree on the VARI-green raster image. The digital values from the VARI-green raster ranged from −101 to 96.

Normalized Difference Vegetation Index (NDVI)
A GreenSeeker handheld crop sensor (Trimble Inc., Sunnyvale, CA, USA) was used to measure the Normalized Difference Vegetation Index (NDVI) for each tree located in the three selected plots 1, 17, and 24. These plots were chosen because of the highest number of dead trees were recorded during the EIM.

Confusion Matrix and Kappa Coefficient
In this study, classification of tree health class was developed based on unsupervised classification available in ArcGIS software. After that, confusion matrix and kappa coefficient were derived for classification. Kappa used to calculate proposed classification method performance, which higher kappa value approaching +1.0 showed strong correlation between classification and validation or reference class data. Furthermore, note that Kappa is not a measure of accuracy but of agreement beyond chance, and thus chance correction is not needed [41].
The confusion matrix was created based on the individual class accuracy that was calculated by dividing the element for each class (row and column). Next, the Producers Accuracy can be generated by dividing each of the major diagonal class with total number of elements on the column of category. The User Accuracy was computed by dividing each of the major diagonal with total number of rows of the element that was classified. Then, the value of kappa coefficient can be generated by using equations below, were the nearest the kappa value to 1, represent the perfect agreement [42].

Tree Health Data and VARI-Green Indices
In Block 42H (aged 1 year 7 months [19 months] old ), a total of 11,069 E. pellita trees were recorded and assigned to one of the four classifications of tree health Table 5 based on infestation levels. Seven trees were classified as dead (VARI-green index −2 to 0), five trees were severely infected (VARI-green index 0 to 0.05), 967 trees were mildly infected (VARI-green index 0.06 to 0.16), and 10,090 trees were considered healthy (VARI-green index 0.17 to 2.00). The individual overall tree health status shown in the map of Block 42H in (Figure 4).
The VARI-green analysis undertaken in the 6-month-old stand (Block 42G) was, however, unsuccessful due to the size of the tree crowns being too small for detection.
Results of 24 plots assessed in Block 42H during early inventory monitoring, groundtruthing, and after VARI-green analysis are given in Figure 5. In total, 1065 trees were recorded as living during the EIM, and 754 trees were recorded as living after VARI-green analysis. In total, 25 trees were recorded dead during the EIM and one was recorded dead after VARI-green analysis. In total, 23 trees were recorded as missing during the EIM, while 255 trees were recorded missing after VARI-green analysis. The summary ground-truth results Table 6, verified one tree as dead (Class 1) and one tree as severely infected (Class 2), respectively, while 58 trees were recorded as being mildly infected (Class 3) and 686 trees were recorded as healthy (Class 4).
The NDVI data (Supplementary Tables S1-S3) were collected to further validate the VARI-green indices data. The resulting correlation between NDVI and VARI-green showed slight differences by plot with a mean correlation value for the three trial plots of~0.73. The highest correlation value was observed in Plot 17 followed by Plot 24 and Plot 1 with values of 0.78, 0.72, and 0.69 respectively.
To assess the accuracy classification of tree health, we adopted a confusion matrix of the classification and kappa coefficient. The confusion matrix in Table 7 shows model prediction for each class are approximately between 60% to 100%, however class II of severe class showed very low accuracy with 0.02%. For this matrix, producer's accuracy was between 71% and 100%, meanwhile user's accuracies were between 0.0% and 98%. The overall accuracy of 0.91 succeeded in identifying almost all tree health categories with more than 90% accuracy.   RI-green analysis. In total, 25 trees were recorded dead during the EIM and one was recorded dead after VARI-green analysis. In total, 23 trees were recorded as missing during the EIM, while 255 trees were recorded missing after VARI-green analysis. The summary ground-truth results Table 6, verified one tree as dead (Class 1) and one tree as severely infected (Class 2), respectively, while 58 trees were recorded as being mildly infected (Class 3) and 686 trees were recorded as healthy (Class 4).

Range of Index Map of VARI-Green as Benchmark for Detection of Health Status Using UAV
In this study, four classifications of health status were developed as a benchmark for detection of individual tree health ( Figure 6). Classification was based on the percentage of red within the VARI-green result for each individual tree. The resulting VARI-green index values were assigned to one of the four health classes.

Range of Index Map of VARI-Green as Benchmark for Detection of Health Status Using UAV
In this study, four classifications of health status were developed as a benchmark for detection of individual tree health ( Figure 6). Classification was based on the percentage of red within the VARI-green result for each individual tree. The resulting VARI-green index values were assigned to one of the four health classes.  Figure 7 shows the procedure to generate VARI-green plots from the raw data. Ortho-mosaic rasters (Figure 7a) from RGB processed data clearly show the normal visualization of selected plot, from which the VARI-green raster is calculated (Section 2.5). The  Figure 7 shows the procedure to generate VARI-green plots from the raw data. Orthomosaic rasters (Figure 7a) from RGB processed data clearly show the normal visualization of selected plot, from which the VARI-green raster is calculated (Section 2.5). The VARI-green raster ( Figure 7b) is a result that shows the selected plot via hyperspectral visualization. To ease analysis, the tree data were isolated from soil to produce the tree-based VARI-green (Figure 7c). Due to the crowded indices data, normalized VARI-green output (Figure 7d) is produced to normalize the data between two classifications of indices, namely, 2 to 0 and 0 to −2. From these data, the classification of tree health status was created, based on the percentage of red in the VARI-green tree crown (Figure 8).
VARI-green raster (Figure 7b) is a result that shows the selected plot via hyperspectral visualization. To ease analysis, the tree data were isolated from soil to produce the tree-based VARI-green (Figure 7c). Due to the crowded indices data, normalized VA-RI-green output (Figure 7d) is produced to normalize the data between two classifications of indices, namely, 2 to 0 and 0 to −2. From these data, the classification of tree health status was created, based on the percentage of red in the VARI-green tree crown (Figure 8).

VARI-Green Pattern and Tree Health Status
The VARI-green classification map quantifies the amount of greenness in an image and differentiates between infestation levels of individual trees based on the relative amount of red in the green image. Dead trees (Class 1) appeared in the VARI-green raster pattern as totally red (Figure 8a), while severe infection (Class 2), caused in this instance by stem borer, was represented by VARI-green raster signatures with up to 80% red at the center of trees encircled by green (Class 2: Figure 8b). For mild infection (Class 3), trees show distortion in foliage and shoot dieback with VARI-green showing scattered red on individual trees (Class 3: Figure 8c). Healthy trees (Class 4) show no sign of infection on foliage and stems. The VARI-green raster pattern for Class 4 shows complete, or nearly complete, green or individual trees (Class 4: Figure 8d). Based on this variation in VARI-green pattern, the majority of trees in Block 42H are almost in perfect condition or healthy.

VARI-Green Pattern and Tree Health Status
The VARI-green classification map quantifies the amount of greenness in an image and differentiates between infestation levels of individual trees based on the relative amount of red in the green image. Dead trees (Class 1) appeared in the VARI-green raster pattern as totally red (Figure 8a), while severe infection (Class 2), caused in this instance by stem borer, was represented by VARI-green raster signatures with up to 80% red at the center of trees encircled by green (Class 2: Figure 8b). For mild infection (Class 3), trees show distortion in foliage and shoot dieback with VARI-green showing scattered red on individual trees (Class 3: Figure 8c). Healthy trees (Class 4) show no sign of infection on foliage and stems. The VARI-green raster pattern for Class 4 shows complete, or nearly complete, green or individual trees (Class 4: Figure 8d). Based on this variation in VARI-green pattern, the majority of trees in Block 42H are almost in perfect condition or healthy.

Discussion
The results of VARI (Visual Atmospheric Resistance Index) image analysis using the green channel suggest that UAV equipped with RGB cameras can be used to monitor plant health status in plantation forests. In this study, the effectiveness of a UAV equipped with a RGB camera was trialed in a young (19-month-old) stand of plantation-grown Eucalyptus pellita in the wet tropics in Sabah, Borneo, Malaysia. Without the need for more expensive hyperspectral cameras, simple RGB images processed using the VARI-green algorithm were able to assess pest and disease early in the growth stage. The result of this study has shown that the infestation and infection of unhealthy E. pellita in industrial tree plantation could be detected with the VARI-green images obtained from a low-cost drone with a low-cost RGB camera across a large area (22.78 ha). This is because VARI-green has been found to be less sensitive to atmospheric effects than NDVI, allowing a good estimation of vegetation fractions as found by [32,39].
Considering the typical host specificity of pest and disease of E. pellita, the classification of VARI-green images can be categorized into four classes. This classification was based on the interpretation with VARI-green index of ortho-mosaic images and ground truth verification. Confusion matrix was developed to calculate misclassification or error classification according to [43]. The result showed that our proposed methods achieve satisfactory performance, with a kappa coefficient of 0.62 and overall accuracy of 91% [44]. However, this study lacks molecular identification, while ground verification of plant health status classes is based on typical symptoms shown by the pathogen and the presence of insect pests. In the case of the dead, Class 1 E. pellita, the main cause of mortality was the soilborne pathogen, Ralstonia solanacearum. During ground inspection, dead trees exhibited leaf drop for entire the crown, branch dieback, and reduced growth. Ralstonia has been reported as a destructive phytopathogen that infects the xylem tissue leading to tree death [2,45] and has been reported to cause mortality in young E. pellita in Indonesia [39].
Based on ground verification, Class 2 was mainly caused by stem borers Zeuzera coffeae (Cossidae) and Endoclita sp. (Hepialidae) with trees showing symptoms of splitting stems and stunted growth. Attack by stem borer has been reported to interrupt nutrient transportation due to swollen stems that cause breaking of the crown top resulting in poor wood quality [46]. For Class 3, trees were infected by mirid bugs, Helopeltis sp. (Miridae). Helopeltis predominantly attack young Eucalypt leaves and shoots causing lesions, curling and drying of the foliage [47]. In addition to pest infestations, tree plantations may also suffer from water stress that causes leaf shedding however most trees retained younger leaves to cope with water deficit [48].
Overall, the early inventory monitoring results show 4% of seedling death and missing trees within the sample plots (0.96 ha) at 3 months post planting. The VARI-green data obtained 19-months post planting show 34% of seedlings as dead or missing in the sample plots (0.96 ha) from VARI-green analysis of RGB images. As a newly planted block, the trees in Block 42H are vulnerable due to their small size. They face hazards such as hot, dry weather, insect infestation, browsing by larger animals [49] or competition from vigorous weed growth although site preparation requires weed-free establishment according to Sabah Softwoods' standard operating procedure [50]. Despite this, it is accepted as normal in plantation forestry that 5-10% of the seedlings may die from one or other of these causes within the first year [49]. As shown in the results (Table 6), the high number of missing trees after VARI-green analysis compared to EIM is up to 30% may influence by several factor. The VARI-green analysis was performed 19 months post planting compared to the EIM assessment at 3 months post planting. Within this period, numerous changes have occurred in terms of tree mortality, due to any one of several biotic and/or abiotic events, e.g., drought and hot weather and infestation of pest and disease. The difference in assessment of dead and missing trees between the VARI-green result and the EIM assessment may be due to errors in either or both survey methods. The RGB images acquired by UAV may suffer from parallax error that occurs due to changes in the relative position of the object or the observation point, whereas double counting or missing counting during the manual early inventory measurement may result in discrepancies. In this study, the VARI-green analysis and ground-truth verification in three plots were able to reduce the counting error with respect to the EIM. These results agree with those of [30] whereby the use of UAV RGB imagery is more effective for estimation of tree heath compared to visual assessments.
From a viewpoint of managing the plantation forest, VARI-green analysis indicates that only 0.06% of trees were dead and 0.04% of tree we severely infected, while 9% of the trees showed mild inhibition and 90% the trees were healthy in the total area of 22.78 ha. In terms of overall survey area, this study is amongst the largest undertaken for the detection of pest and disease in any crop: [31] covered 0.19 ha of potato crop, [39] assessed 3 ha of E. pellita, [30] 0.14 ha of potato crop, and [18] 5.03 ha of vineyard. The development of a health indices map combined with the practical survey of large areas shows the novelty of the study. The health indices map is easy to use, allowing plantation forest managers to view the health status of commercial forests block using comparatively low-cost technology in a rapid manner. Use of UAV to survey forest health would enable considerably larger areas to be assessed (23 ha in 30 min) compared to ground crews (10 ha per day). Aerial surveys of tree health would allow decision-making on operational procedures such as treatment of the infected area or replanting of missing trees.
Detection of early plant stress [51]; nutrient status [52]; and plant phenotyping including height, flower, and canopy cover [53] use NDVI for mapping vegetation condition and status [54]. However, NDVI requires a multispectral sensor operating in the red-edge nearinfrared (NIR) wavelength range [16,17]. Near infrared cameras are more expensive than RGB cameras and require time-consuming calibration procedures [54] unlike VARI-green. VARI-green is based simply on RGB images [18] and can operate at very high resolution using low-cost UAV. This study developed plant health indices using a GoPro consumer RGB camera which was considerably cheaper than a UAV equipped with a hyperspectral camera. Though the reliability and simplicity of VARI-green has earned its popular use [54], other plant indices have been developed using conventional RGB camera. Ref. [32] found VARI-green to be minimally sensitive to atmospheric effects allowing estimation of vegetation fraction with an error of <10% in a wide range of atmospheric optical densities. This study has verified the remotely sensed VARI-green indices with proximal NDVI measures, which have shown acceptable agreement.
The total flight time of 30 min used in this study was shown to be sufficient to cover the survey area of around 23 hectares. Due to the short duration of the flight and the ease of development of plant health indices for E. pellita, it is proposed that UAVs with simple RGB cameras have the potential and cost effectiveness to operate in large-scale monitoring of plant health not only in Malaysia, but also within the tropical countries where E. pellita is grown. The study provides a methodology to assess infection on the ground with aid of remote sensing UAV image for minimal cost. It is an innovative technique that can be included in plantation management plans, not only for tree plantations, but it may also be suitable for other plantation crops such as oil palm. It is not intended that this method be a total replacement of conventional ground verification or conventional surveying methods. Instead, it is intended that aerial survey provide complementary data to rapidly and safely distinguish unhealthy areas of plantation with minimal labor cost. This would allow growers to quickly target affected areas which would be ground-truthed by trained field crews who could back-up the identification of the disease outbreak using molecular identification and morphology identification. The relationship between VARIgreen and ground survey showed good agreement providing a reliable classification tool for plantation managers. The four classes of plant health status have been selected for the management to provide extra information regarding the tree's health enabling better decision-making to control pest and disease outbreaks.
While this study was proven to be reliable for E. pellita at a stand age of 19 months, younger trees failed to be detected by the VARI-green algorithm due to the small crown size. In a forest block of 6-month-old E. pellita (Block 42G), VARI-green analysis was unable to differentiate between trees, indicating that there must be a minimum canopy size before VARI-green can produce reliable result. These findings also suggest that forest plantation deployed with species such as Albizia sp. and Tectona grandis may encounter similar problems of confounding values in VARI-green analysis due to overlapping crown.

Conclusions
There are competing needs for speed and ease of monitoring large areas of planted forest and the need for accuracy and precision. In a practical operation for a forest grower managing tens of thousands of hectares of planted forest, aerial survey of forest health using VARI-green from RGB imaging, is a qualitative assessment that is provides information across a large area at the expense of lower accuracy. It is the relative area that is affected which is important. To determine whether 30% of an area is potentially affected does not require a high degree of accuracy: it is simply enough to know that between 25% and 35% is affected, or less than 10% or more than 50%, etc. The commercial leverage achieved by using UAV remote sensing systems to assess tree health is the ability to survey large areas in a short period of time, and even more powerfully the ability to survey remote and steep terrain safely, without the need for ground crews to be present. This study has shown that significant plantation areas (in this instance 22.78 ha.) can be surveyed in half a day.