What question did this study set out to answer?

The study aims to enhance rice pest detection accuracy in complex agricultural environments using a lightweight model.

March 6, 2026Open Access

HDA-YOLO: a hierarchical and densely-fused attention network for rice pest detection in complex agricultural environments

SYShuo YuanHenan University YDYing DuanHenan University HSHongting SuHenan University

Key Points

The study aims to enhance rice pest detection accuracy in complex agricultural environments using a lightweight model.
Developed HDA-YOLO, a lightweight YOLOv8 model using hierarchical and densely-fused attention mechanisms.
Incorporated asymmetric dynamic downsampling and a multi-scale cascade pre-fusion module for feature fidelity.
Constructed HADF-Net with intra-scale and inter-scale attention modules for content-aware feature fusion.
Upgraded to a multi-scale context module for better adaptability to target scale variations.
Evaluated performance on the RicePest_12 dataset comparing metrics like mAP@50, F1-score, and Recall.
HDA-YOLO reported a 2.4% increase in mAP@50, 3.8% in F1-score, and 4.8% in Recall compared to YOLOv8n.
Achieved a 4.8 percentage point improvement in mAP@50 over the RT-DETR-R18 model with lower computational costs.
Maintained a lightweight structure with 3.93M parameters and 12.02 GFLOPs.

Abstract

Rapid and intelligent identification of rice pests serves as the core sensing technology for precision plant protection and smart rice farming systems, providing critical support for intelligent cultivation decisions. To address the challenges of insufficient robustness and low precision of existing lightweight detection models in complex agricultural environments, this study proposes HDA-YOLO, an improved lightweight YOLOv8 model based on a hierarchical and densely-fused attention mechanism, for fast and high-precision pest detection. To enhance feature fidelity, the model incorporates asymmetric dynamic downsampling (ADDS) and a multi-scale cascade pre-fusion (MCPF) module into the backbone network. To achieve dynamic, content-aware feature fusion, a hierarchical attention-driven dense fusion network (HADF-Net) is constructed, integrating an intra-scale self-attention module (ISAM) and an inter-scale cross-attention module (ICAM). Furthermore, the C2f module is upgraded to a multi-scale context (MSC) module to improve adaptability to variations in target scale. Experimental results on the self-built RicePest₁2 dataset demonstrate that HDA-YOLO, while maintaining a lightweight architecture (3. 93M parameters, 12. 02 GFLOPs), achieves significant improvements over the baseline YOLOv8n model, with mAP@50, F1-score, and Recall increasing by 2. 4%, 3. 8%, and 4. 8%, respectively. In comparison with the Transformer-based RT-DETR-R18 model, HDA-YOLO achieves a 4. 8 percentage points higher mAP@50, while its computational cost is only 22% and its parameter count is only 20% of RT-DETR-R18. Moreover, the proposed model has been successfully deployed on a mobile application, achieving real-time and accurate identification of field pests and demonstrating significant potential in the field of smart rice agriculture.

Perguntar à IA

Bookmark

View Full Paper