Towards Generalizable Crop-Agnostic Plant Disease Recognition: An Unsupervised-Learning Approach

Abstract

In recent years, machine learning techniques have been incorporated in the agricultural industry to automate plant disease recognition. However, most existing frameworks are constrained to the recognition of certain species’ diseases, such as those of tomatoes, due to the severe imbalance in the published dataset. These methods tend to exhibit a strong bias towards specific plants, which require extensive retraining if the model were to accurately classify the disease of other plants. To ensure the system functions in a more realistic context, where images of various plants would be given, it is crucial to develop a generalized system that can recognize a diverse set of plants. To address the aforementioned problem, in this research paper, I propose a novel crop-agnostic plant detection framework. The method leverages contrastive learning, a type of unsupervised learning, to extract discriminative features from plant images regardless of their species. Moreover, the generalizable solution is capable of successfully distinguishing diseases from a dataset with imbalanced class and category. The method achieved an accuracy of 85.88% in the plant village dataset. Through experimental results, it has been demonstrated that the proposed method outperforms existing state-of-the-art methods by a significant margin.