Multi-Disease Classification in In-Field Coffee Leaves Using Cost-Efficient Ensemble Classifier

Computer vision techniques have been instrumental in predicting plant diseases so that preventive measures can be taken on time. The current study uses deep learning ensemble techniques to focus on multiple disease classifications - healthy, red spider mite, and rust in coffee leaves. Transfer learning has been employed using eight pre-trained models - VGG16, VGG19, InceptionV3, ResNet50V2, EfficientNetB0, MobileNetV2, DenseNet201, and InceptionResNetV2, for classifying coffee leaf images into the three classes. An in-field dataset has been used, and data augmentation techniques are employed to overcome the problem of data imbalance. The models have been trained, validated, and tested using a 60-20-20 train-val-test split ratio, and the classification accuracy and the training duration of the models are examined. Analysing the trade-off between the two factors, the top two models have been chosen as base classifiers to constitute a more robust and cost-efficient ensemble classifier. The results show that EfficientNetB0 architecture is the most efficient, with a classification accuracy of 92.48% and a training duration of 1.09 hours. MobileNetV2 architecture is the second top model with 91.16% accuracy and 0.79 hours of training duration. The 2-model ensemble classifier performs with an average accuracy of 97.01% on five random test sets with a total training duration of 1.88 hours. The model also outperforms the state-of-the-art and performs exceptionally well (99.4% accuracy and training time- 1.05 hours) on the Arabica coffee dataset with class labels - healthy, miner, cerscospora, leaf rust, and phoma. Therefore, the proposed model can be used for multi-disease classification in real-time coffee leaves to minimise yield losses.