An ensemble of deep learning architectures for accurate plant disease classification

A substantial fraction of agricultural produce loss can be attributed to plant diseases; agricultural yield loss can have far-reaching consequences for a country's economy and contribute to global food insecurity. Early detection of plant diseases can be instrumental in maintaining global health and welfare. A pathologist's visual evaluation is typically used to make an early diagnosis of plant diseases. This technique involves experts or farmers examining plants with the naked eye and classifying the disease depending on their previous experience. This conventional approach includes drawbacks like low accuracy and the need for human expertise. This motivates researchers to investigate automated systems for the early diagnosis of plant diseases. To achieve this goal an ensemble of different deep learning architectures (DenseNet201, efficientNetB0, inceptionresnetV2, efficientNetB3) is introduced to increase the classification accuracy of plant leaf diseases. In this work, a novel image-processing technique is proposed to increase the efficiency of deep-learning models. Also, a data balancing technique is used to solve the problem of the imbalanced dataset. Five different deep-learning models are trained and tested using the largest plant disease dataset; PlantVillage. Ten different ensembles (chosen randomly) of the deep learning models are tested and compared to find the ensemble with the highest accuracy. The proposed ensemble model was able to achieve 99.89% accuracy on the New PlantVillage dataset. PlantVillage is a challenging dataset with 38 classes. Achieving high accuracies on such a dataset proves the ability of the system to generalize on unseen data or real-world scenarios. A comparison with the state-of-the-art is made with other available models from the literature. A section about this is added to show the superior performance of the proposed ensemble model in terms of accuracy and F1-score.