DIGITAL DISEASE PHENOTYPING Cristhian Delgado , Hernan Benitez * Pontificia Universidad Javeriana Cali, Colombia Maribel Cruz, Michael Selvaraj † International Center for Tropical Agriculture Cali, Colombia ABSTRACT Precise and rapid methods of plant disease detection and evaluation are key factors to accelerate resistant variety development in the rice breeding program. Conventional methods for the disease detection and evaluation is mainly carried out using standard visual estimation by trained experts which is slow and prone to high level of subjec- tivity. Rigorous research has recently recognized innovative, sensor- based methods for the detection and evaluation of plant diseases. Among different type of sensors, aerial multispectral imaging pro- vides a fast and nondestructive way of scanning plants in diseased regions and has been used by various researchers to classify symp- tom levels on the spectral profile of a plant. In this paper, we devel- oped machine learning models to classify rice breeding lines infected by rice Hoja Blanca virus (RHBV) using multispectral images col- lected from UAV (unmanned aerial vehicle). Our results revealed that, the Support Vector Machine (SVM) and Random Forest (RF) methods were not significantly different in their ability to separate susceptible from non-susceptible classes, but SVM best classifiers showed a better sensitivity rates 0.74 (SVM) versus 0.71 (RF). The tool developed from this study will allow rice breeders to charac- terize Hoja Blanca virus resistant varieties considerably earlier, and subsequent in reduced costs. Index Terms— Machine Learning (ML), Unmaned Aerial Ve- hicle (UAV), Rice Hoja Blanca Virus (RHBV), Hight Troughtput Phenotyping (HTP) 1. INTRODUCTION Agriculture researchers require to analyze many traits of thousands of plants across the crop growth cycle. The need to evaluating dif- ferent environments and to carry out replicated trials makes more severe the problem of assessing multiple traits on large populations. [1]. Phenotyping has driven passionate efforts by the plant biologists and engineers to implement innovative machinery in field phenotyp- ing. An instance of that is high-throughput image-based phenotyp- ing, which has unlocked novel projections for non-destructive field- based plant phenotyping for many plant traits, comprising physio- logical, biotic (pest and diseases) and abiotic (heat, drought, flood- ing, nutrient deficiencies) stress traits [2]. Rice Hoja Blanca virus (white leaf) is one of the most intri- cate, baffling, inspiring and remarkable pathosystems come upon on the field of plant virology [3]. The disease is triggered by the rice Hoja Blanca virus (RHBV), which is transmitted by the planthop- per vector Tagosodes orizicolus. The symptoms in the rice plants are chlorotic streaks that can merge and cause the leaves to turn yel- low or white. When young plants are diseased, they are stunted, and in severe infections, the leaves turn necrotic, and finally, the plants * hdbenitez@javerianacali.edu.co † m.selvaraj@cgiar.org die [4]. Besides its important spatial propagation capacity, the main apprehension of agricultural scientists and rice growers was the sub- stantial yield losses that this disease could induce in susceptible rice cultivars, as a result of reduced photosynthetic capacity, plant dwarf- ing, and grain sterility [5]. The development of plant resistance to both viruses and insects has been an effective strategy to combat this disease. The Interna- tional Center for Tropical Agriculture (CIAT) and the Latin Amer- ican Fund for Irrigated Rice (FLAR) are developing resistant rice germplasm to RHBV. This process requires the reliability testing of the new rice genotypes under the influence of the biotic stress for RHBV. CIAT evaluates an average of 20,000 genotypes yearly from national and international plant breeding programs. The evaluation of the phenotypes is carried out in the field, under artificial infes- tation to guarantee that the insect transmits the virus on the plants under evaluation. Given the vast scale of the evaluation process, it serves as screening to discard sensitive materials. The remaining ones, that is to say, those that appear to be resistant to the disease, are evaluated in more specific tests and under controlled greenhouse conditions. The evaluation of affected and unaffected plants is done on a visual scale with three main disadvantages: the bias of the eval- uator, the lengthy time, and the elevated cost. Current approaches to image-based high throughput phenotyping such as the use of aerial images acquired by UAV are objective, fast and affordable alterna- tives to standard phenotyping. Recently, different automatic classification methods have been used to classify remote-sensing data and plant observations. Artifi- cial neural networks and SVM are powerful machine learning tech- niques for supervised classification in plant disease research such as in image-based plant disease detection [6], recognition of plant diseases by leaf images [7] and identification of rice leaf blast us- ing remote-sensing imagery [8]. Previous studies using image-based technologies in the evaluation of RHBV, under field conditions, are not known. 2. MATERIALS AND METHODS 2.1. Unmanned Aerial Vehicle (UAV) and Camera The aerial platform in this project consists of a DJI S1000 1 pro- fessional octocopter (see figure 1a). The A2 autopilot system was programmed for autonomous navigation using PC Ground Station 2 through a data link of 2.4Ghz. PC Ground Station contains software tools such as Photogrammetry, F channel controller, and General Purpose servo action configuration that permitted to plan the UAV path, and configure the PWM output channel and ports to trigger the Camera during the flight. The UAV was equipped with a multispectral camera RedEdge 3 1 https://www.dji.com/spreading-wings-s1000 2 https://www.dji.com/pc-ground-station 3 https://www.micasense.com