Hybrid deep segmentation architecture using dual attention U-Net and Mask-RCNN for accurate detection of pests, diseases, and weeds in crops

Abstract

Early and accurate identification of pests, diseases, and weeds in modern agriculture is crucial for sustainable crop management and yield optimization to increase productivity. This research proposes a hybrid deep segmentation framework that integrates Dual Attention UNet and Mask-RCNN methods to enhance the precision and reliability of plant disease detection under diverse environmental conditions. The core objective is to improve segmentation accuracy and object localization, particularly in complex field imagery with overlapping foliage, variable lighting, and background noise. The proposed architecture uses the Plant-Village dataset, which includes a diverse collection of annotated crop images representing multiple classes of pests, diseases, and weed species. The Dual Attention UNet emphasizes salient spatial and channel-wise features, enabling refined pixel-level segmentation of affected regions. This is followed by a Mask-RCNN module that performs instance-aware segmentation and bounding box localization, facilitating detailed identification of individual anomalies even in cluttered scenes. The framework is further enhanced through data augmentation and transfer learning strategies to support generalization across varying crop types. Experimental evaluation reveals that the proposed deep learning-based model achieves a Detection Accuracy (DA) of 96.5%, an F1-Score of 95.2%, AUC-PR of 97.4%, Sensitivity of 96.5%, Scalability of 96.2% and a Processing Time (PT) of 12 seconds per batch, demonstrating both precision and efficiency. Moreover, the architecture shows a Scalability of 96.8%, ensuring robustness in large-scale deployments. The comprehensive results are compared with baseline models such as CNN, Faster R-CNN, and CBAM. The hybrid integration of instance-aware detection and attention-driven segmentation, explicitly designed for agricultural situations, shows the novelty, and the model improves detection quality by capturing fine-grained spatial characteristics and allowing for the thorough separation of overlapping anomalies compared to traditional CNN-YOLO pipelines. This model presents a reliable solution for real-time smart agriculture systems aimed at proactive crop health management.