DBA-DeepLab: Dual-Backbone Attention-Enhanced DeepLab V3+ Model for Plant Disease Segmentation.

Abstract

Accurate and efficient plant disease segmentation is crucial for early diagnosis and precision agriculture. In this study, we propose a DBA-DeepLab model, i.e., a Dual-Backbone Attention-Enhanced DeepLab model, which integrates DeepLabV3+ with dual backbones of ResNet-50 and EfficientNet-B3 and a Convolutional Block Attention Module (CBAM) for improved plant disease segmentation. The integration of multi-scale feature extraction, attention mechanisms, and edge preservation with the Sobel filter enhances the ability of the model to focus on disease-affected regions with more accuracy and reduce false positives and false negatives. The model was trained and validated using the PlantDoc dataset with a batch size of 32, Adam optimizer, and 50 epochs for better convergence and generalization. Experimental results show that the proposed DBA-DeepLab outperforms DeepLabV3+ with EfficientNet-B3 encoder, DeepLabV3+ with ResNet-50 encoder, and DeepLabV3+ with dual encoder (EfficientNet-B3 and ResNet-50) in terms of segmentation parameters. The proposed model yields 99.35% accuracy, a 91.48% Dice coefficient, an 85.85% IoU coefficient, 96.78% precision, and 100% recall, outperforming the state-of-the-art. Grad-CAM visualization was applied to validate the model's interpretability, affirming its capacity to highlight disease-affected regions and avoid background noise. Comparative analyses with these DeepLabV3+ variants support the improved generalization, segmentation accuracy, and robustness of the proposed model. These results show that DBA-DeepLab is an extremely efficient and scalable solution for plant disease segmentation, with potential applications in smart farming, automatic disease detection, and precision agriculture.

Code and data availability

The paper's segmentation experiments were trained/validated on a public Kaggle leaf disease segmentation dataset (images with ground-truth masks), explicitly declared openly available in the Data Availability Statement. No author analysis code or trained model checkpoints are disclosed.