Plant Disease Recognition Using Transfer Learning and Evolutionary Algorithms

This study used pre-trained convolutional neural network models to perform plant disease recognition. The pre-trained models were fine-tuned and trained on a Plant Village dataset (a publically opened dataset). The dataset contains 54,303 images divided into 38 classes with 14 distinct plant species and maj orly divided into healthy and diseased plants. We implemented different parameters to improve the network's performance during training, including batch size, image size, various numbers of epochs, and class weight. In this study, we used class weight due to the data imbalance (some classes were better represented than others); the reweighting technique tested different image sizes to evaluate the performance. The pre-trained models were used for transfer learning by freezing the network's last layer. The models used in this study are MobileNetV2, EfficientNet-B5, and InceptionV3; they acquired a diseased classification accuracy of 96.53%, 98.73%, and 96.39%, respectively. The transfer learning models' predictions were combined using two types of ensemble techniques; grid search ensemble and differential evolution algorithm, with improved accuracy of 99.14% for grid search algorithm and 99.16% for differential evolution algorithm in classifying diseased and healthy plants. The ensemble method's successful classification of diseased and healthy plants is promising. It can positively impact the further improvement of plant disease recognition and serve as an early warning tool for farmers and consultants in real-time conditions.