Hierarchical attention and feature enhancement network for multi-scale small targets in pine wilt disease

Pine Wilt Disease (PWD) is a forest disease caused by the pine wood nematode, which poses a serious threat to the health of pine trees. This disease leads to the death of pine trees and disrupts the stability and biodiversity of the forest ecosystem. Early detection and prevention can maximize the success rate of treating pitch canker disease and minimize its impact on the environment. Therefore, early detection and control of pitch canker disease is of utmost importance. However, existing models often struggle to distinguish the early infection characteristics from the noise in the complex background environment, resulting in a high rate of false detections. In response to these challenges, our research introduces HAFENet, a robust detection model tailored for small PWD-infected targets. HAFENet incorporates a Two-Stage Attention Fusion Module, designed to effectively extract target features in complex environments. This module adopts a self-attention cross-mechanism to fuse the global features and local features processed by the weight learning module, thereby paying more attention to small target regions and improving the model’s accuracy in early small target detection tasks. Additionally, to prevent redundant features in the environment from obscuring the subtle early infection characteristics, we designed an Information Integrity Convolution (IIConv). This dual-branch structure processes redundant and important features separately and then merges them using a stacking technique, which suppresses irrelevant information interference; Furthermore, HAFENet introduces a DeCoupled Head to separately optimize the classification and localization heads. We utilize Normalized Wasserstein Distance to provide more effective localization error feedback for detection boxes. Validation on our established small-target PWD dataset shows that HAFENet achieves an accuracy of 86.0% for early-stage detection and 96.7% for late-stage detection, representing improvements of 4.3% and 1.1% over the baseline, respectively. Compared with the existing mainstream models, HAFENet achieves the highest accuracy in the detection of small targets at the early stage of pine wood nematode infection. Additionally, it exhibits strong anti-interference capabilities on test sets with added noise and blur. These results indicate that HAFENet can maintain robust performance even in harsh environments with noisy backgrounds, highlighting its potential for wide application in forest protection and management.