Automated Guava Disease Detection Using Transfer Learning With ResNet‐101

The old method of identifying diseases, which farmers used by manually inspecting their farms, is ineffective because it is time-consuming, prone to human error, and cannot be applied to a wide agricultural territory. Sustainable agriculture thus requires automated disease detection that provides accurate results to meet the increasing demand for this technology. This paper explores the use of deep learning (DL) and transfer learning (TL) using ResNet-101 to advance the guava disease detection. The image information is raw, and the image information is directly given to ResNet-101, which can recognize complex patterns without manually extracting features and hence creating an effective and accurate classification method. To make the sample size bigger and more balanced, data augmentation was applied to the original collection of 3784 images and resulted in the generation of 4632 images in equal proportions in three categories of health condition: Anthracnose, Fruit Fly, and Healthy guavas. The balanced dataset was separated into three parts: the training, the validation, and the test parts that consisted of 80%, 10%, and 10%, respectively, to make sure that the model is well trained and tested. The preprocessing of the data was also done by normalization and resizing methods, which improved the performance of the model. The accuracy of the proposed model was found to be impressive, with 98.48% being the percentage of accuracy in the classification of the guava disease, and it is exhaustively tested in conserving nine main measures, which are accuracy; misclassification rate, specificity, recall, precision, negative predictive value (NPV), false positive rate (FPR), false negative rate (FNR), and F1 score. In order to provide interpretability, fairness, and transparency, Gradient-weighted Class Activation Mapping (Grad-CAM) visualization was used, generating heatmaps that reveal the diseased areas and which also make sure that the network concentrates on the real areas of infection. It is an artificial intelligence (AI) technology that provides a better identification of plant diseases and is a viable solution in large-scale agriculture.