A Weighted Voting Deep Learning Approach for Plant Disease Classification

Abstract

Plant destruction is usually caused by plant diseases. Without early and reliable detection, it can affect the plants and may eventually cause permanent losses, especially for inexperienced farmers. Therefore, intelligence in agriculture is becoming more and more required. Thereafter early diagnosis and classification are crucial and a very challenging research field in the agriculture sector for its treatment. In this context, many solutions have been proposed. Deep learning has been highly successful and hardly applicable in this problem. However, through pass survey analysis, we notice that there are a few studies in DL for disease classification problems but the precision and outcomes of different traditional DL methods may vary and give a less score for classification. The proposed approach is based on a weighted combination of five deep learning architectures. The weight of each DL architecture is calculated based on its performance using genetic algorithms. The results of the proposed approach are evaluated on the publicly available Plant Village (PV) dataset. It is found that using the Deep Learning weighted voting method gives higher classification accuracy compared to the results obtained using each DL architecture separately and also compared to recent approaches in literature, which allowed us to correctly identify the leaves and to improve the classification accuracy rate to 99.21%.