An Incremental Learning for Plant Disease classification

Artificial neural networks suffer from catastrophic forgetting unlike humans when these networks are trained on something new, which causes significant performance degradation. This study presents the application of incremental learning for plant disease classification. As farmers need to know the newly originate disease in their farms, the study of incremental learning is important in agriculture areas who want to know the knowledge of newly born disease as time goes on. To do this, we took 36 classes of the PlantVillage dataset and randomly split them into six equal incremental groups. The two versions of ResNet called ResNet18 and ResNet50 are used to train six groups of classes iteratively. The test accuracy of the classes is measured in each incremental step. The first group of incremental classes achieves 100% accuracy in the test data, whereas, the second group achieves only 57% accuracy due to imbalanced data between the classes. The highest accuracy for third, fourth, fifth, and sixth is 82.7%, 93.4%, 92.9%, and 90.5% using ResNet50. We hope that incremental learning for plant disease classification helps to identify the newly born plant diseases in a sequence of time.