Detectando a ferrugem asiática na folha da soja utilizando redes neurais convolucionais

Abstract

Soybean cultivation is the main agricultural crop in Brazil. The main disease responsible for reaching this crop and decreasing the production potential of the plant in Brazil is the Asian soybean rust. This work presents a solution of a mobile application for the Android platform that is able to identify the Asian soybean rust when it is present in the leaf of the plant using the camera of an Android device, in this way, helping the rural producers to carry out the treatment of the disease in the soybean plantations as soon as possible. The work was developed applying computational vision techniques and a Convolutional Neural Network model to perform image classification. The TensorFlow library was used throughout the entire application development process from the artificial neural network training process to the real-time identification of the disease in the soybean leaf. Inception-V3 Convolutional Neural Network model was used to perform in-app image classifications, a famous model known for having great results in Deep Learning problems and image classifications. In total, 2770 images of different leaves of soybean were collected for the development of the work, being divided between images for performing neural network training and imaging tests for application validation. At the end of the tests, the developed application reached an accuracy of 84.61% in the detection of Asian rust on soybean leaf. At the end of the tests, the trained model was shown to be efficient in the separation of healthy soybean leaflets and Asian soybean rust, some difficulties were also observed in distinguishing other diseases from the soybean crop, even with the results obtained in satisfactory.