PWD-lightweight and feature fusion network for multi-stage joint detection of pine wilt disease

Pine Wilt Disease (PWD) has caused irreversible damage to the health of pine forests around the world. Accurate detection and identification is the prerequisite for taking measures to prevent the spread of PWD. To tackle this challenge, this study presents PWD-Lightweight and Feature Fusion Network (PWD-LWFFNet), a specially designed object detection model for detecting PWD in pine trees. PWD-LWFFNet uses lightweight EIBNet as its backbone network, which is mainly implemented by the EIB module for lightweighting. The EIB module utilizes the inverted bottleneck module and incorporates Extra Depthwise convolution (DW) to minimize the number of parameters while ensuring computational efficiency. In the neck network, PAFPN-4Net was designed as a multi-scale feature fusion network and an FMN module was introduced. The FMN module merges the input features with the global information through its ‘aggregation’ and ‘modulation’ components. This setup allows the network to dynamically focus its attention on the minute disease details in the early stages of pine tree infection. Four detection heads are designed and integrated with the Enhanced Multi-scale Attention (EMA) mechanism to capture fine-grained features. Finally, PIoUv2 is selected as the loss function to guide anchor boxes along the optimal path for regression. Comprehensive experiments demonstrate that PWD-LWFFNet exhibits excellent performance in detecting PWD, which the mean Average Precision (mAP) is 94.1%. It particularly excels in detecting small, early-stage targets compared to other mainstream models, with an Average Precision (AP) of 83.4%. The detection accuracies for the middle, late, and tree mortality stages reach 97.3%, 98.4%, and 97.2% respectively. When compared with existing mainstream models, PWD-LWFFNet demonstrates state-of-the-art performance. Experiments conducted on the PWD dataset established in this paper show that PWD-LWFFNet maintains good performance even in environments with background noise, validating the effectiveness of the model in early disease detection and management in pine forests. Its lightweight design provides a guarantee for practical application deployment.