Modified transfer learning frameworks to identify potato leaf diseases

Abstract

Potato diseases such as early and late blight are the most lethal diseases that can cause significant damage to potato production. Detecting these diseases early and making a precise diagnosis will help the farmers to manage the crop effectively and prevent disease spread. Due to its high accuracy, deep learning, specifically the Deep Convolutional Neural Network (DCNN) models, have gained popularity in image identification and classification. In this research, three transfer learning (TL) pre-trained models, namely VGG19, NASNetMobile, and DensNet169, were modified for detecting potato leaf disease. These models were selected based on their proven robust performance in various computer vision tasks. The performance of these pre-trained models was improved by introducing additional layers to their original architecture, thereby reducing trainable parameters. Furthermore, three state-of-the-art TL models, ResNet50V2, InceptionV3, and Xception, were also trained to detect potato disease and their performance compared with our three modified models. The experiment was conducted on a dataset collected from Kaggle, which was then divided into training, test, and validation sets. The performance of these models was tested with a confusion matrix, Matthew's correlation coefficient (MCC), Cohen's kappa coefficient (CKC), Area Under the Curve (AUC), and ROC (Receiver Operating Characteristics) curve. The modified DenseNet achieved an accuracy of 99 percent, MCC of 98.5 percent, CKC of 98.5 percent, and an AUC-ROC score of 0.990 for the test set, while an accuracy of 100 percent, MCC of 99.5 percent, CKC of 99.5 percent, and AUC-ROC score of 0.997 for the validation set. The modified NasNet and VGG19 achieved an accuracy of 98 percent and 94 percent, MCC of 97 percent and 90.7 percent, and CKC of 97 percent and 90.5 percent for the test set, while an accuracy of 95 percent and 96 percent, MCC of 92.7 percent and 93.6 percent, CKC of 92.5 percent and 93.5 percent for the validation set. The five fold cross-validation method was applied, and overall results showed that the modified DenseNet model outperformed the state-of-the-art TL models.