DWP-YOLOv11: A Lightweight Object Detection Network for UAV Imagery Under Complex Weather Conditions

Abstract

Object detection in UAV imagery under complex weather conditions remains challenging due to weather-induced image degradation, limited target visibility, and densely distributed small objects. To overcome these limitations, a lightweight detection framework, termed DWP-YOLOv11, is developed based on the YOLOv11 architecture. First, the backbone network is made lightweight by replacing standard convolutions with depthwise separable convolution (DWConv), thereby reducing computational overhead while maintaining effective feature extraction. Second, a Dual-Path Multi-Scale Spatial Pyramid Fusion (DMSPPF) module is introduced to combine complementary max-pooling and average-pooling information, improving multi-scale feature representation for low-texture and low-contrast targets. Third, the existing Parallel Position-Aware Attention (PPA) module is incorporated to strengthen discriminative feature learning and reduce background interference caused by complex weather conditions. The detection head is also redesigned by incorporating a high-resolution P2 detection branch and eliminating the unnecessary P5 branch, thereby strengthening the detection capability for small- and medium-scale objects. Comprehensive experiments on the self-constructed UAV Complex Weather Dataset (UCWD) indicate that DWP-YOLOv11 attains 72.2% mAP@0.5 and 50.9% mAP@0.5:0.95, corresponding to improvements of 11.8 and 9.2 percentage points over YOLOv11s, respectively. Moreover, DWP-YOLOv11 surpasses HA-YOLO and ADWNet while requiring only 8.7 M parameters and 20.2 G FLOPs, providing an effective balance between predictive accuracy and computational cost.