Convolutional neural networks for the recognition of diseases and pests in Cassava leaves (Manihot esculenta)

Abstract

In this work, we recognized cassava diseases and pests, by means of convolutional neural networks, as a way to avoid the spread of pathogens, prevent economic losses, and favor decision-making for a proper disease management. For the development of this system, VGG16, ResNet50 and Xception models were chosen for having displayed good performance in previous researches of disease classification in plants, which we considered very similar to our case of study. For the training procedure, a transfer learning technique was implemented, employing a database categorized by cassava diseases (bacterial blight, brown streak, green mite, mosaic disease), as well as healthy leaves. This database was balanced and refined manually, selecting the images that represented characteristics of each category, according to the description found in the existing literature. Finally, the best model was chosen taking into account its performance measured through the Accuracy metric. The best model obtained, which we propose in this work, was Xception, and was trained during a period of 35 epochs with 6120 images of cassava leaves, achieving an accuracy of 94.56% . This model provides an option to detect cassava leaf diseases early, reliably and cost-effectively.