Leaf spot is a devastating disease in cultivated peanut (Arachis hypogaea L.) that can lead to significant yield losses without chemical controls. Multiple disease symptoms, two causal organisms, inconsistent testing environments, and genotype-by-environment interactions are all components which make breeding for leaf spot resistant peanuts challenging. To better understand this disease, and make gains in breeding for disease resistance, an accurate and objective phenotyping strategy must be implemented. In this work, data derived from leaf scans, UAV-captured RGB, and multispectral imagery were evaluated as a replacement for the subjective visual rating scale used at present. Standard operating procedures are detailed for all digital methods evaluated in this paper, and all digital phenotypes are fully characterized with descriptive statistics. Feature importance and post hoc proof of concept studies are conducted to further evaluate the new digital methods. Ultimately, ‘Visible Atmospherically Resistant Index’ or VARI was selected as the most appropriate proxy for visual ratings and should be deployed by researchers and plant breeders in the peanut community for the objective evaluation of leaf spot resistance.

Germplasm

A population of 264 allotetraploid peanut genotypes were selected for this study because of their value to the North Carolina State University (NCSU) Peanut Breeding and Genetics (PB&G) program. The population consists of 12 NCSU PB&G cultivars, 165 breeding lines, 85 lines with wild species introgressions, and 2 plant introduction lines (Supplemental Table 1). By choosing advanced and highly related genotypes, any findings from this study can be quickly incorporated into PB&G breeding activities. Moreover, the 264 peanut lines represent a gradient of resistant to susceptible phenotypes in response to leaf spot disease pressure.

Field Design

The population of 264 genotypes was evaluated at the Peanut Belt Research Station (PBRS, Lewiston-Woodville, NC) and the Upper Coastal Plains Research Station (UCPRS, Rocky Mount, NC), over the summers of 2021 and 2022; creating 4 unique combinations of year and location, hereafter referred to as experiments. Each experiment had two replications and each replication was planted in a randomized complete block design. Each genotype was planted as a two-row plot with the following dimensions: 90 cm row centers and 7.3 meter row length. An alleyway of 2.1 meters separated the ranges. Besides the absence of fungicides, fields were managed with standard cultural practices for the region. To facilitate georeferencing, within 30 days after planting, 6 to 9 ground control points were installed in the field in a uniform manner. Ground control points were constructed of Polyvinyl Chloride (PVC) pipe of 5.08 cm diameter and 100 cm length with a Polyvinyl Chloride 5.08 cm cap. The pipe was driven into the soil until the cap was flush with the ground surface. An Emlid Reach RS2 (Emlid, Hong Kong, China) Real-Time Kinematic (RTK) receiver was used to measure the x, y, and z coordinates of each ground control point.

Phenotyping Pipelines

The phenotypes presented in this study were collected when peanuts were in growth stage R8 (Subrahmanyam et al., 1995). For the calculation of defoliation only, the UAV was flown at R7 to capture change relative to the R8 stage. Visual ratings, leaf scans and UAV data were all captured within 48 hours of each other. Early and Late Leaf Spot were rated as one composite disease.

Visual Ratings

All visual ratings were collected by the same experienced rater. Visual ratings were collected on leaf spot symptoms using the modified 9-point scale as presented by (Subrahmanyam et al., 1995) in every plot of every experiment.

Low Throughput Leaf Scans

For every genotype, eight leaves (equaling 32 leaflets) were randomly selected from non-border vegetation. Leaves were placed in whirl packs from Weber Scientific (Hamilton, NJ) and refrigerated until scanning. Leaves were scanned within seven days of field collection to avoid desiccation and discoloration. Leaflets were arranged so that no overlapping occurred and the adaxial side of the leaf was scanned in color with an Epson DS-50000 flatbed Scanner (Los Alamitos, CA). Lesion count and lesion area on a plot-basis were calculated in Python version 3.6.10 using Supplemental Script 1.

UAV-Captured RGB Images

Before each flight, 15 cm square PVC targets that were colored in a black and white checker pattern, were centered over all RTK-surveyed ground control points present in the field. A DJI (Shenzhen, China) Matrice 300 RTK UAV was flown with a DJI Zenmuse P1 35mm lens to capture true color Red Green Blue (RGB) images. The UAV was flown at an altitude of 30 meters, with a speed of 7.8 miles per hour (ground sample distance of 0.38 cm/px). The DJI Pilot flight plan ensured 80/80 image overlap to facilitate image stitching. Images from each flight were imported into Metashape version 1.6 (St. Petersburg, Russia). Using the default workflow, images were aligned, a dense point cloud and digital surface model were built, and the orthomosaic was created. The known RTK positions of the ground control points were imported and used to geometrically rectify the orthomosaic product. Orthomosaics were exported into the EPSG:2264 projected coordinate system and imported into ArcGIS Pro (Esri, Redlands, California) version 2.3.0 for plot segmentation, plot labeling, removing soil pixels, calculation of percentage defoliation, and extracting reflectance values on a plot-basis (Supplemental Note 1).

UAV-Captured Multispectral Images

In a separate flight from the RGB image collection, the same DJI Matrice 300 RTK UAV was flown with a 10-band RedEdge-MX dual sensor from Micasense (Seattle, WA). Figure 1 shows the wavelength location and width captured by each band. Immediately before each UAV flight, a Micasense RP06 calibrated reflectance panel was imaged with the RedEdge-P dual sensor. The UAV was flown at an altitude of 40 meters, with a speed of 7.8 miles per hour (ground sample distance of 2.73 cm/px). The DJI Pilot flight plan ensured 80/80 image overlap to facilitate image stitching. Metashape version 1.6 was used as in section 2.2.3, with the following modification; images were color calibrated to produce an absolute surface reflectance using the preflight image of the calibrated reflectance panel prior to image alignment. ArcGIS Pro version 2.3.0 was used for plot segmentation, plot labeling, removing soil pixels and extracting reflectance values on a plot-basis (Supplemental Note 1). In Python version 3.6.10 with statsmodels version 0.13.2, analysis of variance was used to determine if any of the nine classes of visual ratings were significantly different at the ten wavelengths recorded by the multispectral camera.

Multispectral and RGB based Vegetation Indices

Digital Numbers (DNs) representing reflectance were calculated and summarized on a plot mean basis from the UAV imagery (section 2.2.3 and 2.2.4). These values were read into Python version 3.6.10 for Vegetation Index (VI) calculation (Supplemental Script 2). Table 1 lists all VIs calculated and the associated bands used (Blackburn, 1998; Broge & Leblanc, 2001; Chappelle et al., 1992; Daughtry et al., 2000; Elvidge & Chen, 1995; Gitelson et al., 2002; Gitelson & Merzlyak, 1994; Haboudane et al., 2004; Huete et al., 2002; Huete, 1988; Jordan, 1969).