Classifying Unbalanced Potato Plant Diseases using Ensemble Deep Learning Models

Diseases have always been considered one of the major challenges for the efficient and profitable growing of potatoes, with such diseases as Early blight and late blight causing high yield loss and financial costs all around the world. Even though deep learning algorithms proved to be quite successful in the task of automated plant diseases identification, their efficiency is often hampered due to class imbalance in the datasets, lack of annotated agricultural datasets and low performance in practice, due to bad generalization from training to real-world situations. Besides, most existing researches concentrate on the issue of overall accuracy, ignoring the question of class-wise accuracy, especially for minority classes of the diseases that are important for practical disease diagnosis. That is why there is a necessity in developing a robust classification system that will be able to deal with class imbalance in the datasets and give reliable and accurate disease identification.

In the current paper, an ensemble deep learning framework for potato leaf disease classification using imbalanced dataset is proposed. The methodology includes the following steps: Image preprocessing, data augmentation, transfer learning, ensemble learning and recommendation of a fertilizer. The images of potato leaves were divided into three classes: Healthy, Early Blight and Late Blight. Four deep learning classifiers-Custom CNN, ResNet50, VGG16 and MobileNetV2 were used as the base learners using transfer learning when possible. The predictions of the models were combined using Hard Voting, Soft Voting and Stacking ensembles. The evaluation of the framework was conducted with the help of Accuracy, Precision, Recall, F1-score, ROC-AUC, 5-fold cross-validation, confidence intervals and class-wise performance. Also, the agronomic decision-making module for fertilizer recommendation was included in the system.

As shown by experiments, transfer learning models perform much better than the Custom CNN base model, while ensemble learning improves the classification accuracy even more. Among the tested methods, the proposed Stacking ensemble gives the best results with the accuracy of 99.70%, precision of 99.70%, recall of 99.70% and F1-score of 99.70%. It also shows the smallest variability of the predictions (standard deviation 0.0012) and narrowest 95% confidence interval (0.9955-0.9985). Besides, the proposed method demonstrates superior class-wise performance, which confirms its efficiency for dealing with the problem of class imbalance in agricultural datasets.