Optimizing Rose Leaf Disease Detection Performance via Explainable Transfer Learning: A Comparative Analysis

This paper advocates a novel method for identifying disorders in rose leaves by applying transfer learning techniques. The methodology includes data collecting and preparation steps like reshaping, gray scaling, resizing, and augmentation. Initially, a basic Convolutional Neural Network (CNN) is utilized, followed by pretrained transfer learning techniques on ImageNet. Specifically, DenseNet201, EfficientNetB3, InceptionResNetV2, and MobileNetV2 are used. The analysis is performed based on categorical losses, recall, accuracy, and precision. It is evident from comparing the experimental results to the CNN model's baseline that transfer learning greatly increases accuracy, precision, and recall. Applying CNN transfer learning models to the DenseNet201 architecture produces the best results, with accuracy of 94.41%, precision of 96.18%, recall of 92.77%, and validation accuracy of 98.91%. The outcome of DenseNet201is also visually analyzed through layer wise explainability. Results show the effectiveness of transfer learning and how its improving the CNN model's ability to diagnose diseases in rose leaves. Proposed study particularly promotes automated disease identification in plants in the field of rose leaf diseases. The findings offer valuable information for managing plant health and agricultural activities in future.