MSDF-Net: a cross-version lightweight detection framework based on deformable convolution and high-resolution feature enhancement for pine wilt disease.

Introduction: Early and precise identification of pine wilt disease is critical for effective control. However, early-stage lesions are extremely small, sparse, and scattered, making them highly susceptible to being obscured by noise in complex forest backgrounds. Moreover, feature simplification in lightweight deployment further leads to the loss of critical pathological information. Methods: We propose MSDF-Net, a lightweight object detection framework that integrates a high-resolution P2 detection layer for enhanced small-target sensitivity, DCNv4-based deformable convolution for adaptive modeling of irregular spatial patterns, an EMA attention mechanism for background suppression, and a dual-branch C2f DualConv module for efficient multi-scale feature fusion. The model was evaluated on a cross-regional dataset spanning three provinces and two pine species. Results: MSDF-Net achieves an mAP@0.5 of 80.1%, outperforming the YOLOv8n baseline by 5.1 percentage points while maintaining 2.67M parameters and 11.7 GFLOPs. The most substantial improvement occurs in early-stage disease detection (PWD-E), with an AP gain of 20.2 percentage points. Cross-version validation on YOLOv11n, YOLOv12n, and YOLOv13n yields consistent improvements of 6.1, 4.3, and 5.1 percentage points in mAP@0.5, respectively. Discussion: Given its effectiveness across multiple YOLO versions and ecological conditions, MSDF-Net provides a generalizable solution with low parameter count and moderate computational complexity, making it a promising candidate for future UAV edge deployment pending on-device validation.