Transfer Learning and Data Augmentation Based CNN Model for Potato Late Blight Disease Detection

Abstract

Plant disease management is an essential step in the process of detecting pathogens in plants. For diseases like Potato's Late Blight, ineffective management could destroy the whole farm within a day. As a result, the total yield per unit of the potato becomes diminished. In this paper, potato's late blight disease detection model was built using the CNN algorithm. The dataset is collected in two ways. The first is preparing a dataset by capturing an image of the leaf from the Holeta potato farm, and the other is using the benchmark dataset. The dataset has two classes: the first class has a healthy class category and the other Late Blight. One of the problems with machine learning is not having enough data. In our case, to train a model publicly available database images of 596 and 430 of our own images were used. To address the problem of a small dataset we have used data augmentation techniques and transfer learning along with 5-fold cross-validation. InceptionV3, VGG16, and VGG19 pretrained models were used for transfer learning techniques. InceptionV3 model achieved 87% score among other pretrained models while testing with unseen data. In the future, the performance of the model could be improved by having a sufficient amount of dataset. Convolutional Neural Network Deep learning Plant disease detection Pretrained model Potato's Late Blight Convolutional Neural Network Deep learning Plant disease detection Pretrained model Potato's Late Blight