high throughput phenotyping of tomato spot wilt disease in ......disease detection were determined...

9
4 IEEE Instrumentation & Measurement Magazine June 2017 1094-6969/17/$25.00©2017IEEE High Throughput Phenotyping of Tomato Spot Wilt Disease in Peanuts Using Unmanned Aerial Systems and Multispectral Imaging Aaron Patrick, Sara Pelham, Albert Culbreath, C. Corely Holbrook, Ignácio José de Godoy, and Changying Li T he amount of visible and near infrared light re- flected by plants varies depending on their health. In this study, multispectral images were acquired by a quadcopter for high throughput phenotyping of tomato spot wilt disease resistance among twenty genotypes of peanuts. The plants were visually assessed to acquire ground truth ratings of disease incidence. Multispectral images were pro- cessed into several vegetation indices. The vegetation index image of each plot has a unique distribution of pixel intensi- ties. The percentage and number of pixels above and below varying thresholds were extracted. These features were cor- related with manually acquired data to develop a model for assessing the percentage of each plot diseased. Ultimately, the best vegetation indices and pixel distribution feature for disease detection were determined and correlated with man- ual ratings and yield. The relative resistance of each genotype was then compared. Image-based disease ratings effectively ranked genotype resistance as early as 93 days from seeding. Introduction The supply of food, fiber, fuel, and feed needs to be doubled to meet the demand of the rapidly growing global population that is projected to rise over 9 billion by the middle of the cen- tury [1]. Meanwhile, global climate change will make it more difficult to maintain, much less increase, the current yields of major crops. Higher temperatures are expected to increase the prevalence of pests and diseases in major crops [2]. Peanuts are a staple food for many countries and are susceptible to a vi- ral disease known as tomato spotted wilt (TSW). The tomato spotted wilt virus (TSWV), a tosporvirus in the family Bun- vaviridae, causes the disease and was first discovered to affect tomatoes in 1915 [3]. Although it is named after its first known host, there are more than fourteen distinct tospoviruses in the family of viruses that afflict a multitude of crops. The virus is estimated to have caused millions of dollars in loss of peanut production in the United States over the past few decades [4]. In response to the devastation by the virus, resistant genotypes of peanuts have been bred. The use of disease resistant crops reduces the labor and resource demands and increases yields. The current standard method for assessing disease resistance or susceptibility requires the tallying of foot long segments of a plot that are affected [5]. While this method is effective, it is time and labor intensive and therefore, impractical for very large studies. The intensities of visible and non-visible light reflected by plants can be used to determine the health of a plant as a whole or detect diseased areas of a leaf. Near infrared reflectance measured by a handheld fiber optic probe has been used to screen sugarcane for disease and pest resistance for breeding purposes [6], and vegetation indices (VIs) have been proven effective in segmenting and classifying disease in rice [7]. Unmanned aerial systems (UAS) are increasingly used in agri- culture due to their ability to collect data with high spatial and temporal resolution [8]. Spectral analysis similar to the stud- ies mentioned above has been performed on data collected by aircraft for the purpose of disease detection. Hyperspec- tral imaging has been used to map mosaic virus in sugarcane [9] and identify stripe rust in winter wheat [10]. Multispectral imaging has been proven effective in detecting Huanglong- bing in citrus trees [11] and laurel wilt disease in avocados [12]. Multispectral data was used to estimate yield in rice [13] and compare the emergence percentages of winter wheat va- rieties [14]. No studies have reported the detection of spotted wilt in peanuts using multispectral imaging and UAS. In ad- dition, previous studies detect disease and classify the state of disease into two or three categories with the intent of treat- ment of the disease, and the studies do not account for genetic variety. There remains a need for quantification of disease re- sistance and susceptibility across genetic varieties using data collected by UAS. To fill these knowledge gaps, the overall goal of this study was to investigate the feasibility of aerial multispectral imaging in high throughput phenotyping of TSW resistance in peanuts. Specifically, the objectives were to: identify the best vegetation index for the detection and quantification of spotted wilt; explore the possibility of predicting crop yield using the identified vegetation index;

Upload: others

Post on 09-Mar-2021

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

4 IEEE Instrumentation & Measurement Magazine June 20171094-6969/17/$25.00©2017IEEE

High Throughput Phenotyping of Tomato Spot Wilt Disease in

Peanuts Using Unmanned Aerial Systems and Multispectral Imaging

Aaron Patrick, Sara Pelham, Albert Culbreath, C. Corely Holbrook, Ignácio José de Godoy, and Changying Li

T he amount of visible and near infrared light re-flected by plants varies depending on their health. In this study, multispectral images were acquired by a

quadcopter for high throughput phenotyping of tomato spot wilt disease resistance among twenty genotypes of peanuts. The plants were visually assessed to acquire ground truth ratings of disease incidence. Multispectral images were pro-cessed into several vegetation indices. The vegetation index image of each plot has a unique distribution of pixel intensi-ties. The percentage and number of pixels above and below varying thresholds were extracted. These features were cor-related with manually acquired data to develop a model for assessing the percentage of each plot diseased. Ultimately, the best vegetation indices and pixel distribution feature for disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype was then compared. Image-based disease ratings effectively ranked genotype resistance as early as 93 days from seeding.

IntroductionThe supply of food, fiber, fuel, and feed needs to be doubled to meet the demand of the rapidly growing global population that is projected to rise over 9 billion by the middle of the cen-tury [1]. Meanwhile, global climate change will make it more difficult to maintain, much less increase, the current yields of major crops. Higher temperatures are expected to increase the prevalence of pests and diseases in major crops [2]. Peanuts are a staple food for many countries and are susceptible to a vi-ral disease known as tomato spotted wilt (TSW). The tomato spotted wilt virus (TSWV), a tosporvirus in the family Bun-vaviridae, causes the disease and was first discovered to affect tomatoes in 1915 [3]. Although it is named after its first known host, there are more than fourteen distinct tospoviruses in the family of viruses that afflict a multitude of crops. The virus is estimated to have caused millions of dollars in loss of peanut production in the United States over the past few decades [4]. In response to the devastation by the virus, resistant genotypes of peanuts have been bred. The use of disease resistant crops reduces the labor and resource demands and increases yields.

The current standard method for assessing disease resistance or susceptibility requires the tallying of foot long segments of a plot that are affected [5]. While this method is effective, it is time and labor intensive and therefore, impractical for very large studies.

The intensities of visible and non-visible light reflected by plants can be used to determine the health of a plant as a whole or detect diseased areas of a leaf. Near infrared reflectance measured by a handheld fiber optic probe has been used to screen sugarcane for disease and pest resistance for breeding purposes [6], and vegetation indices (VIs) have been proven effective in segmenting and classifying disease in rice [7]. Unmanned aerial systems (UAS) are increasingly used in agri-culture due to their ability to collect data with high spatial and temporal resolution [8]. Spectral analysis similar to the stud-ies mentioned above has been performed on data collected by aircraft for the purpose of disease detection. Hyperspec-tral imaging has been used to map mosaic virus in sugarcane [9] and identify stripe rust in winter wheat [10]. Multispectral imaging has been proven effective in detecting Huanglong-bing in citrus trees [11] and laurel wilt disease in avocados [12]. Multispectral data was used to estimate yield in rice [13] and compare the emergence percentages of winter wheat va-rieties [14]. No studies have reported the detection of spotted wilt in peanuts using multispectral imaging and UAS. In ad-dition, previous studies detect disease and classify the state of disease into two or three categories with the intent of treat-ment of the disease, and the studies do not account for genetic variety. There remains a need for quantification of disease re-sistance and susceptibility across genetic varieties using data collected by UAS.

To fill these knowledge gaps, the overall goal of this study was to investigate the feasibility of aerial multispectral imaging in high throughput phenotyping of TSW resistance in peanuts. Specifically, the objectives were to:

◗ identify the best vegetation index for the detection and quantification of spotted wilt;

◗ explore the possibility of predicting crop yield using the identified vegetation index;

Page 2: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

June 2017 IEEE Instrumentation & Measurement Magazine 5

◗ determine the earliest time at which the disease incidence and relative resistance can be detected; and

◗ infer and rank the resistance across genotypes.

Materials and Methods

Crop Study SetupOn April 27, 2016, we planted eighty plots of peanuts. Each plot consisted of two twenty foot rows. The study utilized twenty genotypes of peanuts with four plot repetitions of each (Fig. 1). Georgia-06 (GA-06G) and Tifguard (a new nematode resistant peanut variety) are two genotypes that have some field resis-tance to TSW [15]. The purpose of the study was to determine whether peanut genotypes from Brazil had any resistance to TSW. To prevent leaf spot and other fungal diseases, we treated the plots with fungicide but did not treat them with insecticide. Thrips are the primary insect vector for TSWV, and application of insecticides would confound the study for TSW resistance.

Manual Data CollectionThroughout the study, we performed manual assessment of disease progression. We determined the proportion of each plot afflicted by segmenting the plot into one-foot sections and recording the number of afflicted segments. This com-monly used assessment method classifies any segment that is stunted, dead, or experiencing chlorosis as afflicted with TSWV [5]. The number of feet of the plot affected ranged from 0 to 40 ft. For simplicity, we converted the ratings to percent-ages to use in the figures and discussion of the data. We also performed immunoassays on root tissue to ensure the pres-ence of TSWV.

Image Data CollectionA custom data acquisition system was created by integrat-ing a MicaSense RedEdge multispectral camera (MicaSense, Seattle, WA, USA) into a DJI Phantom 3 quadcopter (DJI, Shenzhen, China) (Fig. 2). The Phantom's original gimbal

Fig. 1. (a) Genotype names, country of origin. (b) Labeled layout of test field on 6/29/2016.

Page 3: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

6 IEEE Instrumentation & Measurement Magazine June 2017

and camera operate on 12 V, so a switching power regulator was used to attain a 5 V power source for the data acquisition system. The system consists of a microcontroller, 10 degree of freedom sensor, micro SD card, 3D printed bracket, and the original vibration dampeners. The system calculates the angle of the camera using the accelerometers and gyro-scopes and sensor fusion via a Kalman filter. A barometer is used to determine the altitude. Images are collected only when the camera is stable and at the desired height. We used the SD card for recording the pose of the camera to generate 3D models (However, we did not prepare that to include in this paper). A 3DR uBlox GPS module connects directly to the multispectral camera to georeference images. The MicaSense RedEdge camera has five spectral bands: blue (475 nm), green (560 nm), red (668 nm), red edge (717 nm), and near infrared (840 nm). We took photos of a white

calibration pad with the multispectral camera before and af-ter flight image collection for image correction.

Image Processing and Data AnalysisWe developed the following processing pipeline to process aerial multispectral images and vegetation indices and to correlate image features to manual assessments (Fig. 3). After we acquired the images, we uploaded them to the MicaSense ATLAS data processing system. The cloud service aligned and stitched the images together. We downloaded the im-ages as a five-layered GEOTIFF. We cropped, rotated, and split the GEOTIFF into its component channels within MAT-LAB (MathWorks, Natick, MA, USA). We generated an RGB color image and exported it. Using the color image, we made a plot segmentation mask in Adobe Photoshop (Adobe, San Jose, CA, USA). Isolation of the plot required selecting the plots, ensuring separation between them, and avoidance of weeds and border rows. We produced VI images of each plot using the multispectral images and the mask. In total, we se-lected 23 vegetation indices, many of which came directly from a relevant study [12]. Other VIs are well established indices, and a few are variants of well-known indices that substitute the red edge for red. These VI images were used to generate histograms of the VI pixel distributions within each plot. The measurements taken in the field, which determine the percent of plot disease, are of a spatial nature. We used histograms to visualize the data because they showed the distribution of pixels and therefore, can indicate what por-tion of the plot was diseased. Diseased portions of the plot would have higher or lower VI intensities than healthy por-tions, depending on the index used (Fig. 4).

VI intensities can be used for comparison within a data set without extensive calibration and correction for lighting and atmospheric conditions. Creating absolute standards for in-terpreting vegetation index values is difficult due to variations in the lighting and atmospheric conditions, the subject im-aged, and the equipment and processing used. Determining a threshold to separate healthy and diseased plot areas became our challenge, since no established standard exists.

To determine each vegetation index, we used the following process. Using each value in the range of the index image in-tensities as a threshold, we recorded the proportion and count of pixels above and below the thresholds within each plot. Then, we applied linear regression to these variables, with the manual ratings as the response variable. We then recorded the threshold and regression equation that resulted in the best co-efficient of determination.

Once we determined the threshold and regression equa-tion, the manual ratings and VI derived estimated ratings were grouped by genotype and compared. The distributions of manual and image derived ratings were tested for normal-ity using Shapiro Wilk and Kolmogorov-Smirnov tests. Then, we subjected the distributions to a paired T-test to determine which genotypes have a significant difference in means by rating method and do not follow the regression trend. We as-sessed the accuracy of regression models by observing errors

Fig. 2. (a) An exploded view of the data acquisition system. (b) UAS in flight from below.

Page 4: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

June 2017 IEEE Instrumentation & Measurement Magazine 7

generated from a leave one out method. We performed the sta-tistical analyses with MATLAB.

Results and Discussion

Best Vegetation Indices for Disease DetectionOverall, strong correlations were found between several VI threshold derived values and the manual ratings (Table 1). The strength of correlations between index image derived val-ues and the manual assessments of the plots increased as the growing season progressed. When plants were first infected, the diseased areas on the leaves were likely too small to detect at a UAV's effective flight height. As the disease advanced, the infected areas became larger and were detectable due to the compound impact of the disease on canopy size and its health, which influenced the image derived scores.

The majority of VIs can produce strong correlation late in the season, and the best characterizing feature was typically the number of pixels above the threshold. Disease detection seems to imply the detected feature should be the number of diseased pixels, but the better indicator is the number of healthy pixels. Unhealthy plants will have lower VI values (for most VIs) but will also be smaller. Therefore, the better determinant of plot health is the amount of healthy canopy area.

Across the third and fourth data pairs, the NDRE provided the best correlation. The Ratio VI - RE index provided the same results and was identical to the NDRE when linearized with the threshold method. These indices relied on the same bands (RE, IR) and were highly dependent on each other. Similarly,

Fig. 4. Distribution of NDRE intensities in the healthiest, most diseased, and the average of all plots. Image data from 8/25/2016 correlated to manual ratings from 8/23/2016 to determine threshold.

Fig. 3. Data processing pipeline.

Page 5: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

8 IEEE Instrumentation & Measurement Magazine June 2017

the green normalized difference VI and green ratio VI were virtually the same when linearized. Due to their redundancy, they are not considered independent when discussing effi-cacy of VIs.

There was a significant separation between the pixel dis-tributions of healthy and diseased plants (Fig. 4). The best threshold neatly bisected the average NDRE intensity distri-butions of plots with the highest and lowest manual ratings. Applying the theshold to classify diseased and healthy por-tions of plots creates a disease map (Fig. 5).

Yield PredictionCrop yield had a strong correlation with VI pixel counts used for disease detection, manual measurements, and canopy area (Fig. 6). NDRE showed not only the strongest correlation with yield among the four VIs used but also turned out to be a better predictor of yield than manual ratings. The superior relation-ship was likely a result of the error caused by discretizing a plot

into feet and the subjectivity of human ratings. Yield was cor-related with the canopy area, but the correlation was weaker than any other variable listed.

Early Differentiation of Disease ResistanceWe conducted regression analyses to observe how early in the season the relative resistance or susceptibility of each plot could be determined. We correlated the image data with the final manual disease incidence ratings. The resulting scatterplots, coefficients of determination, and error statistics demonstrated the increasing effectiveness as the season progressed (Fig. 7). A standard deviation in error of 12.5% indicated that the 95% confidence interval allows for estimates with ±25% error. This is a reasonable tolerable limit in distribution of the percent er-ror, as it allowed for differentiation between the healthiest and most susceptible 50% of genotypes. The regression from the final image collection (8/25/2016) showed a very strong corre-lation between the manual and image derived ratings and the

Table 1 – Vegetation indices used and their best correlations between manual and image derived ratings. IR and RE represent infrared (717 nm) and red edge (840 nm), respectively. “-RE” is used to indicate equation

modified to use the red edge band in place of red.

Image Data Days from Seed

50 63 93 120

Manual Data Days from Seed

44 54 92 118

VI Name Formula R2

NDRE (IR-RE)/(IR+RE) 0.090 0.426 0.716 0.829

Ratio VI -RE IR/RE 0.089 0.379 0.717 0.829

Normalized Near Infrared -RE IR/(IR+RE+G) 0.033 0.374 0.707 0.827

Green Ratio VI IR/G 0.077 0.356 0.685 0.811

Green Normalized Difference VI (IR-G)/(IR+G) 0.077 0.355 0.685 0.811

Normalized Red -RE RE/(IR+RE+G) 0.081 0.310 0.655 0.798

Difference VI -RE IR-RE 0.056 0.273 0.739 0.795

Normalized Near Infrared IR/(IR+R+G) 0.115 0.303 0.661 0.793

Normalized Green G/(IR+R+G) 0.071 0.248 0.611 0.775

Normalized Green -RE G/(IR+RE+G) 0.072 0.225 0.597 0.753

Green Difference VI IR-G 0.031 0.225 0.693 0.751

Green VI (G-RE)/(G+RE) 0.050 0.161 0.579 0.731

Ratio VI IR/R 0.059 0.228 0.601 0.716

Normalized Difference VI (IR-R)/(IR+R) 0.101 0.178 0.601 0.715

Excess Green Minus Excess Red ExG-1.4*RE-G 0.058 0.021 0.173 0.715

Difference VI IR-R 0.101 0.205 0.649 0.692

Normalized Pigment Chlorophyll Index (RE-B)/(B+RE) 0.064 0.016 0.316 0.681

Normalized Excess Blue 1.4*B-G/1.4*B+G 0.125 0.072 0.583 0.669

Excess Green (ExG) |2*G-RE-B| 0.049 0.134 0.576 0.669

Normalized Red R/(IR+R+G) 0.101 0.209 0.506 0.550

Excess Red 1.4*RE-G 0.061 0.072 0.552 0.529

Normalized Red -RE R/(IR+RE+G) 0.011 0.177 0.488 0.518

Woebbecke Index (G-B)/(RE-G) 0.066 0.143 0.414 0.222

Page 6: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

June 2017 IEEE Instrumentation & Measurement Magazine 9

Fig. 5. (a) Color image of the test site. (b) NDRE image. (c) Image separating plot area above (green pixels represent healthy plants) and below (red pixels represent infected plants) derived threshold (0.4584). Image data from 8/25/2016 correlated to manual ratings from 8/23/2016 to determine threshold.

Fig. 6. Correlation between crop yield and plot disease evaluation by manual ratings and image vegetation indices, as well as canopy cover. MR-Manual Ratings from 8/23/2016 (%) NDRE- normalized difference red edge, NRRE - normalized red (red edge), GDVI - green difference vegetation index, GNDVI- green normalized difference vegetation index. Counts above respective thresholds from 8/25/2016 images, yield in pounds and canopy area in pixels (CAP). Pearson correlation coefficient (R) displayed per correlation.

Page 7: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

10 IEEE Instrumentation & Measurement Magazine June 2017

standard deviation of error of 10.28% was precise enough to separate resistant from susceptible genotypes. The July image set provided slightly weaker correlation but was still accept-able for this purpose. The mid and late June sets had some correlations but had an unacceptably high standard deviation of error for predicting the final state of the disease progression. At the time of the earlier evaluations, plants were still grow-ing rapidly. In addition, for early-season manual assessments, stunting was often the primary symptom on which classifica-tion was based. In some cases, stunted plants that are visible to a rater may have been overgrown by spreading limbs of adja-cent plants and possibly not detectable with remote imaging. Although it would be desirable to be able to discern differ-ences in TSW among genotypes throughout the season with NDRE derived disease estimates, being able to differentiate among genotypes late in the season is most critical for most comparisons, regardless of the method of disease assessment. In situations where large numbers of genotypes are compared, one or two manual evaluations may be all that is practicable. In such cases, evaluations are typically made later in the season to

consider the stunting and foliar spotting symptoms that typi-cally develop earlier in the season with general chlorosis that typically develops later in the season [4].

Genotype RankingThe distributions of disease ratings demonstrated the su-perior disease resistance of Tifguard, GA-06G, and GA-12Y (Fig 8). The resistance of Brazilian genotypes varies, and the NDRE ratings can separate the best from worst. Discrepan-cies can be attributed to varying plant size, pigments, severity of affliction, and the subjectivity of human assessment, and plot segmentation. Even with an experienced evaluator, the manual method used in this study was subject to human er-ror. Whether diseased plants are noticed may be affected by, among other factors, plant growth habit, moisture stress, or the intensity and angle of sunlight at the time of rating. Thus, although manual ratings are the standard method, it cannot be concluded which method is actually more accurate.

The regression method used for modeling the relationship between VI derived values and manual ratings was suitable for

Fig. 7. Correlation plots showing the relationship between NDRE derived percent disease predictions based on the correlation of pixel count above the best fit thresholds and the manual ratings taken on 8/23/2016 (118 days). R2, mean absolute errors, and standard deviation of errors.

Page 8: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

June 2017 IEEE Instrumentation & Measurement Magazine 11

the majority of genotypes used in this study. However, some outliers did not conform to the trend line (GA-12Y, IAC137, IAC322, IAC599). This was likely the result of differing pig-ments in the leaves and naturally varying crown sizes. There are many small sources of error which compound to create the nor-mal distribution of errors. The differences between the manual rating method and image rating method are one source of the discrepancies. The manual rating method attributed TSW pres-ence as diseased largely based on the center stems of plants of the row, even if lateral branches of the plants were not affected. In some cases, the diseased area recorded for manual ratings may have been smaller than the pixel resolution of the images. In addition, the manual method discretized the plot by linear feet, not pixels. The image-derived values were influenced by human error in plot segmentation and the rotation of the image.

ConclusionThis study proved the feasibility of using multispectral im-age data acquired by UAS to high throughput phenotype

TSW disease in peanut plots of varying genotypes. Crop yield correlated strongly with image data. Among many vegetation indices, the data revealed that the NDRE was best suited for these purposes. The number of pixels above an iteratively determined threshold provided excellent correlation to the manually assessed percent of plot area affected. Disease and relative resistance, based on the fi-nal manual assessment, could be determined as early as 93 days into the season. The processed images and manually collected data both indicated that the United States bred re-sistant genotypes of peanuts are superior to the Brazilian genotypes trialed. The use of UAS can save hours of labor and provide more consistent results than visual examina-tion by humans.

References[1] H. C. J. Godfray, J. R. Beddington, I. R. Crute, et al., “Food security:

the challenge of feeding 9 billion people,” Science, vol. 327, pp.

812-818, 2010.

Fig. 8. Boxplot of distributions of manual (8/23/2016) and NDRE derived (8/25/2016) ratings by genotype. Sorted by ascending average manually determined percent diseased. **P-Value of paired T-test less than α=0.05, *** less than α=0.01.

Page 9: High throughput phenotyping of tomato spot wilt disease in ......disease detection were determined and correlated with man-ual ratings and yield. The relative resistance of each genotype

12 IEEE Instrumentation & Measurement Magazine June 2017

[2] P. K. Anderson, A. A. Cunningham, N. G. Patel, F. J. Morales, P. R.

Epstein, and P. Daszak, “Emerging infectious diseases of plants:

pathogen pollution, climate change and agrotechnology drivers,”

Trends in Ecology and Evolution, vol. 19, pp. 535-544, 2004.

[3] D. Peters, “Tospoviruses,” in Virus and Virus-like Diseases of Major

Crops in Developing Countries, G. Loebenstein and G. Thottappilly,

eds., Boston, MA, USA: Kluwer Academic Publishers, pp. 719-

742, 2003.

[4] A. Culbreath, J. Todd, and S. Brown, “Epidemiology and

management of tomato spotted wilt in peanut,” Annual review of

Phytopathology, vol. 41, pp. 53-75, 2003.

[5] A. Culbreath, J. Todd, D. Gorbet, F. Shokes, and H. Pappu, “Field

response of new peanut cultivar UF 91108 to tomato spotted wilt

virus,” Plant Disease, vol. 81, pp. 1410-1415, 1997.

[6] D. Sabatier, C. Moon, T. Mhora, R. Rutherford, and M. Laing,

“Near-infrared reflectance (nir) spectroscopy as a high-

throughput screening tool for pest and disease resistance in a

sugarcane breeding programme,” in Proc. 86th Annual Congress

of the South African Sugar Technologists' Assoc. (SASTA 2013), pp.

101-106, 2014.

[7] S. Phadikar and J. Goswami, “Vegetation indices based

segmentation for automatic classification of brown spot and blast

diseases of rice,” in Proc. 3rd Int. Conf. Recent Advances in Inform.

Technology (RAIT), pp. 284-289, 2016.

[8] Y. Shi, J. A. Thomasson, S. C. Murray, N. A. Pugh, W. L. Rooney,

S. Shafian, et al., “Unmanned aerial vehicles for high-throughput

phenotyping and agronomic research,” PloS one, vol. 11, p.

e0159781, 2016.

[9] É. A. S. Moriya, N. N. Imai, A. M. G. Tommaselli, and G. T.

Miyoshi, “Mapping mosaic virus in sugarcane based on

hyperspectral images,” IEEE J. of Selected Topics in Applied Earth

Observations and Remote Sensing, 2016.

[10] L.-S. Huang, J.-L. Zhao, D.-Y. Zhang, L. Yuan, Y.-Y. Dong, and J.-C.

Zhang, “Identifying and mapping stripe rust in winter wheat

using multi-temporal airborne hyperspectral images,” Int. J.

Agriculture and Biology, vol. 14, 2012.

[11] F. Garcia-Ruiz, S. Sankaran, J. M. Maja, W. S. Lee, J. Rasmussen,

and R. Ehsani, “Comparison of two aerial imaging platforms for

identification of Huanglongbing-infected citrus trees,” Computers

and Electronics in Agriculture, vol. 91, pp. 106-115, 2013.

[12] A. I. de Castro, R. Ehsani, R. C. Ploetz, J. H. Crane, and S.

Buchanon, “Detection of laurel wilt disease in avocado using low

altitude aerial imaging,” PloS one, vol. 10, p. e0124642, 2015.

[13] D. Stroppiana, M. Migliazzi, V. Chiarabini, A. Crema, M. Musanti,

C. Franchino, et al., “Rice yield estimation using multispectral

data from UAV: A preliminary experiment in northern Italy,” in

Proc. 2015 IEEE Int. Geoscience and Remote Sensing Symp. (IGARSS),

pp. 4664-4667, 2015.

[14] S. Sankaran, L. R. Khot, and A. H. Carter, “Field-based crop

phenotyping: multispectral aerial imaging for evaluation of

winter wheat emergence and spring stand,” Computers and

Electronics in Agriculture, vol. 118, pp. 372-379, 2015.

[15] A. Shrestha, R. Srinivasan, S. Sundaraj, A. K. Culbreath, and D. G.

Riley, “Second generation peanut genotypes resistant to thrips-

transmitted Tomato spotted wilt virus exhibit tolerance rather

than true resistance and differentially affect thrips fitness,” J.

Economic Entomology, vol. 106, pp. 587-596, 2013.

Aaron Patrick ([email protected]) is a Research Assistant at the University of Georgia pursuing a master's degree in en-gineering. He received his undergraduate degree from the University of Georgia in agricultural engineering with an emphasis in electronics. His area of research is the use of un-manned aerial systems in high throughput phenotyping.

Sara E. Pelham ([email protected]) is a Research Assistant at the University of Georgia where she will obtain a master's degree in plant pathology. She will pursue a doctorate in crop and soil sciences at the University of Georgia upon completion of her master's degree. She received her undergraduate de-gree from Abraham Baldwin Agricultural College in biology. Her area of research is unmanned aerial systems in agricul-tural research.

Albert Culbreath ([email protected]) is a Professor in the Uni-versity of Georgia, Department of Plant Pathology, located at the Tifton Campus. He received his B.S. (botany) and M.S (plant pathology) degrees from Auburn University and his Ph.D. degree in plant pathology from North Carolina State University. His area of research is epidemiology and man-agement of foliar fungal diseases and tomato spotted wilt of peanut.

Ignácio José de Godoy ([email protected]) is a Scientific Researcher at the Instituto Agronomico, Campinas, SP, Brazil. He received his B.S. degree in agronomy from the University of São Paulo, SP, Brazil and his Ph.D. degree in agronomy (plant breeding) from the University of Florida. His major area of re-search is peanut breeding.

C. Corley Holbrook ([email protected]) is a Research Geneticist with the United States Department of Ag-riculture – Agricultural Research Service, located in Tifton, GA. He received his B.S. (agronomy) and M.S. (plant breeding) degrees from the University of Florida and his Ph.D. degree in plant breeding from North Carolina State University. His area of study is breeding and genetic research on resistance to biotic and abiotic stresses in peanuts.

Changying “Charlie” Li ([email protected]) is Professor in the School of Electrical and Computer Engineering at the College of Engineering of the University of Georgia. He received his Ph.D. degree in biological and agricultural engineering from Pennsylvania State University. His research interest is phe-nomics and plant robotics. He is the recipient of the 2016 New Holland Young Researcher Award from the American Society of Agricultural and Biological Engineers. As the Project Di-rector, Dr. Li currently is leading two large multidisciplinary teams from 10 institutions to develop robotic technologies for high throughput phenotyping of plants.