A Hybrid CNN Ensemble Approach with Feature-Level Fusion for Apple Leaf Disease Classification

The deterioration of apple leaves through disease imposes a tremendous impact on the quality of fruit production from apple orchards. Therefore, swift and accurate identification of these diseases is necessary for effective fruit management. Traditional methods for identifying apple leaf diseases utilize expert humans who inspect the leaves visually. However, visual inspections are labor-intensive, highly subjective, and cannot be conducted effectively in larger orchards. The rapid advancement of deep learning research is facilitating automated identification of apple leaf diseases from images of contaminated leaves using Convolutional Neural Networks (CNNs) that are able to learn visual characteristics such as spots, lesions, colour changes, or texture differences on apple tree leaves. However, when using a single CNN model to accomplish this task, one may encounter difficulties due to environmental variability, similarities between classes of diseases, and generalisation across variable field conditions. The current work proposes a hybrid ensemble CNN model for the analysis of apple leaf disease by providing a solution to these challenges. The ensemble model consists of EfficientNet-B0, DenseNet121, and ResNet50. Feature-level fusion of these models concurrently enables the extraction of strong and complementary feature representations from images for improved performance. The proposed hybrid CNN ensemble model achieved a classification accuracy of 99% when evaluated on a dataset of apple leaf disease images. Furthermore, the proposed model consistently outperformed backbone models such as EfficientNet-B0, VGG16, ResNet50, and MobileNetV2. Therefore, this proposed model has the potential to be an effective, reliable, scalable method to perform disease monitoring in apple production systems.

Cited via DOI — this article is permanently archived on Zenodo.