Classificação de distúrbios em folhas de macieiras utilizando redes neurais convolucionais

Abstract

Over the past few years in Apple culture, losses of up to 25% of production have been verified, these losses are related to several factors, such as climate change, the lack of innovative and competitive technologies for the apple sector, in addition to the appearance of new pests and/or diseases and agravation of losses for reasons already identified. Many of the technological solutions to solve or prevent these problems is the fast recognition in the field of the causes of these losses, especially when there is the incidence of diseases, pests and nutritional imbalances. Nevertheless, the existing method for identifying symptoms in apple rely on using the knowledge of experts to detect symptoms in orchards, using as aid publications containing pictures and descriptions of possible symptoms to be found. Thus, it shows a need for a new approach to addressing this problem on a less costly and more efficient manner. Machine learning techniques have become popular because they can learn information and patterns from data and then predict new results or classify new data submitted. Some of the techniques that are currently state of the art in image classification are the Convolutional Neural Networks, along with artificial neural networks and Support Vector Machines. As an poorly explored area in the literature, it showed the need to create a new image data set. In a way to analyse if the mentioned machine learning techniques can correctly classify the symptoms on leaves. Thus, this thesis proposes the use and comparison of these machine learning techniques to the classification task of symptoms of 2 types of diseases, two symptoms of nutritional deficiencies and one symptom of damage by herbicide in apple leaves. After the establishment of training, validation and test sets, these were used to train and test the techniques of Convolutional Neural Networks, Multilayer Perceptron Networks and Support Vector Machines. In addition, for better comparison of the results obtained, the test sets were sent to agronomy experts for identification of symptoms, in order to compare the accuracy on the results obtained. The results showed that the approach proposed in this paper, using Convolutional Neural Networks , obtained accuracies between 93.3 % and 97.3 %, matching or obtaining superior results to those obtained by other machine learning techniques, as well as by selected experts. Thus, the use of Convolutional Neural Networks may facilitate diagnosis in leaves of apple in quickly, accurately and usual manner.