Efficient Diagnosis of Coffee Leaf Diseases Using Pre-Trained ResNet50 Neural Networks

The method used in this study for automatically detecting the diseases in coffee leaves is the deep learning method based on the pre-trained ResNet50 model. They have used a dataset of images of coffee leaves having four classes, namely miner, no disease, phoma and rust. The steps included in the data preprocessing were resizing of the images to ensure that the model receives input images of equal dimension and augmenting the dataset in a bid to enhance its generalization. ResNet50 model was fine-tuned to have the ability of down streaming features specific to coffee diseases after being pre-trained on ImageNet dataset. This model was developed with an additional dense layer with activation, and the Adam optimizer for the model's training. Evaluation parameters and training metrics depicted that the developed model is highly accurate in classification and an early stopping has been implemented to shy away from over-fitting of the model. To measure the efficiency of the proposed model, a confusion matrix and a classification report have been employed and the accuracy, recall, and F1-score are presented for different classes. Furthermore, the model was converted into TensorFlow Lite Model as it was possible to implement it for real time illness detection on the edge node. Essentially, the results prove that the ResNet50 model is capable of classifying the ills affecting the coffee leaf and this can make it easier for farmers to manage the diseases as they arise in agricultural practices.