Visão computacional aplicada à avaliação de Hemileia vastatrix em Coffea arabica

Abstract

The main disease that affects coffee is orange rust, caused by the fungus Hemileia vastatrix. As a result, the focus of coffee breeding programs is to obtain genotypes resistant to the pathogen, to minimize the damage caused by it to coffee yield. The main strategy adopted for the evaluation of disease-resistant genotypes is the use of standar area diagrams, using notes regarding the reaction of the genotype in relation to the severity of the disease. However, such evaluation are made visually and hard depends of the experience of the evaluators.An alternative for assessing the severity of the disease is the use of photographic images and processing them in a software to obtain the real severity and obtain more assertive and conclusive results. In view of this, the presente work aimed to develop an algorithm for quantifying orange rust and, based on the analyzes obtained through this, to develop and validate a new standard area diagram for evaluating coffee rust. Two experiments were conducted, in wich the first consisted of collecting coffee leaves affected by the disease for training and obtainin na algorithm model for quantification, while the second was carried out on detached leaves inoculated with the fungus, placed on Petri dishes, in a controlled enviroment conditions, with the proposal to develop a new method of inoculation end evaluation of the disease. The comercial varieties used in this study were Catuai Vermelho IAC 144, Bourbon Amarelo, MGS Aranãs, MGS Paraíso and Catiguá MG2. Image processing and analysis were perfomed in python programming language, using the OpenCV and Scikit-Image package. To elaborate the. standard area diagram, the images and estimates severities obtained by the image analysis were used. The developed standard area diagram was validated using the concordance correlation coeficiente proposed by Lin (1989) and proved to be effective in quantifying diseased coffee leaves, while the developed algorithm was also assertive in relation to the disease quantification. Visual assessment methods were compared in person and remotly and no significant difference was identified between the assessments.