Enhanced Attention Residual U-Net and Attention Enhanced EfficientNetB3 Model for Multiclass Plant Leaf Disease Detection
Year: 2025
Type: article
Abstract: Leaf viruses cause large financial losses and represent a serious threat to the world's crop safety.Accurate segmentation and classification of these diseases are essential for maintaining agricultural productivity.Deep Learning (DL) frameworks have been employed to detect leaf infections, reducing the reliance on crop experts.Among several models, the Cascading Autoencoder with Attention Residual U-Net (CAAR-UNet) has shown promise in segmenting and classifying leaf diseases by combining attention and residual connections.However, in leaf images where disease pixels are sparsely distributed throughout the leaf, the normal attention method used in CAAR-UNet might not be able to detect all infected patches.A new Cascading Autoencoder and Enhanced Attention-based Residual U-Net (CAEAR-UNet) model is proposed in this paper to solve above issue.This model utilizes enhanced attention strategy that leverages bit-level information of pixels from leaf images for accurate segmentation of diseased regions.Then, feature extraction is executed by a stacked model with pre-trained Convolutional Neural Networks (CNNs) such as Xception, InceptionV3, and InceptionResNetV2.Disease classification is then carried out using an EfficientNetB3 model with a spatial attention module to focus on disease-related information and classify multiple categories of leaf diseases.The integration of spatial attention module in EfficientNetB3 identifies the disease regions location in leaves and identifies their correlations across multiple diseased regions that enhances the classification accuracy for multiclass leaf disease detection.Hence, the proposed model, named Enhanced Attention Residual U-Net and Attention-enhanced EfficientNetB3 (EARUAE), enhances multiclass leaf disease segmentation and classification accuracy by leveraging correlations among diseased areas.Finally, the experimental results illustrate that the segmentation performance of CAEAR-UNet achieves the highest values for metrics (Mean Pixel Accuracy (MPA) of 0.9794, Weighted Mean Intersection over Union (WMIoU) of 0.8027, a Weighted Mean Boundary F1-score (WMBF1) of 0.6855 and Weighted Mean Dice Coefficient (WMDC) of 0.6149) than CAAR-UNet.In addition, disease classification accuracy 99.98%, precision 99.95% and recall 99.98% of EARUAE is higher than Dual Attention Binary Neural Network (DABNN) and channel Extension Residual with Channel attention Plant leaf detection Network (ERCP-Net) for customized PlantVillage Datasets.
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W4407928453