Transfer learning for classification of bean plant leaf diseases in precision agriculture

Abstract

Bean plant is one of the most valuable cash crops worldwide. However, it is susceptible to serious diseases hindering its production and therefore, early detection and diagnosis is crucial. This study used a dataset contain- ing healthy and infected bean leaf images. It had 1035, 133 and 128 images on train, validation and test sets respectively. Image data augmentation was performed on the dataset to artificially increase the training set, by creatingdifferent versions of leaves. This was done to expose the classier to a massive amount of data during training and ensure model robustness even in case of leaf angle/ light variations. The study adopted transfer learning approach using pre-trained MobileNet model and obtained 4 models from its architecture; MobileNet model without/ with data augmentation and fine-tuned MobileNet model without/ with data augmentation. The models were com- pared and fine-tuned MobileNet model with data augmentation produced the best generalization accuracy of 95% on validation and 94% on test sets. To validate the results obtained from the best classifier, sensitivity analysis was performed by i. training our best model on 10 random samples from the data. Onesample Student T - Test was used to examine the statistical difference between sample mean accuracy and our known best ac- curacy. At 5% level of significance, this difference was found to be statistically insignificant. ii. Training our best classier using different random seeds and hyper-parameters (learning rates, mini-batch sizes & epochs). Similarly, there was a small difference in accuracies obtained, proving model robust- ness. A 95% confidence interval was constructed to provide a range of values within which our overall mean could lie. The interval was [94.94, 96.19], im- plying that there was a 95% chance that the confidence interval contained the true overall mean accuracy. The robust network recorded a very low misclassification count of 7 on validation set and 8 on test set. Model interpretability results revealed that, there was a misclassification only when the model was not paying attention to the leaf or due to odd cases like presence of a hole on the leaf otherwise, there was a correct prediction. Therefore, towards precision agriculture, the robust system developed can be successfully employed in bean disease diagnosis even in case of leaf angle/ light variations.