ESE-PWDNet: an efficient early-stage pine wilt disease detection network

To address the challenge of difficult small target recognition in the early detection of Pine Wilt Disease (PWD), this study proposes an efficient Unmanned Aerial Vehicle remote sensing detection model named ESE-PWDNet (Efficient Small-scale Early PWD Detection Network). Using a DJI Air3 UAV platform, a multi-temporal and multi-view high-resolution dataset of early-stage PWD was independently constructed in the Tangshan Forest Area, Jiangning District, Nanjing City, Jiangsu Province. Based on this dataset, a key module-the Efficient Visual Linear Unit (EFVLU)-was designed. Serving as the foundational building block of ESE-PWDNet, the EFVLU combines with convolutional modules (Conv) to form the backbone network, which efficiently captures global dependencies and improves the detection of small targets through global context while reducing computational complexity. Furthermore, inspired by the PANet architecture and utilizing the Attention State Space Block (ASSB), a novel neck network was designed to empower the model with efficient high-resolution image processing capabilities while maintaining high computational efficiency. In the prediction head, the introduction of the Efficient Multi-scale Attention (EMA) mechanism and the Lightweight Shared Detail Enhanced Convolutional Detection Head (LSDECD) comprehensively enhances the model's perception and localization capabilities for small targets with almost no additional inference computational cost. Experiments on the constructed multi-environment early PWD dataset demonstrate that ESE-PWDNet significantly improves the recognition performance of tiny disease targets in complex scenes. The final model maintains high inference efficiency, achieving a Precision (P) of 75.9% and a Recall (R) of 75.1%, with a low computational complexity of 6.5 GFLOPs and 2.6M parameters. Its comprehensive performance outperforms mainstream comparative models. This research provides a reliable technical solution and data foundation for the early and precise UAV remote sensing monitoring of forestry pests.