Lightweight detection of grapevine downy mildew using UAV based on image information enhancement and adaptive dual-path fusion attention.

Abstract

BACKGROUND: Early detection of grapevine downy mildew (GDM) using unmanned aerial vehicle (UAV) imagery remains highly challenging due to limited onboard computing power, natural light overexposure, limited imaging resolution that obscures subtle early lesions, pronounced symptom differences between leaf surfaces, and severe canopy occlusion. MODELS: To address these challenges, a lightweight GDM detection network (GDM-Net) for UAV-based applications is proposed. A CLAHE–Unsharp Masking–Gamma Correction (CUG) image enhancement framework is proposed to achieve synergistic multi-dimensional feature enhancement by expanding the regional dynamic range, emphasizing high-frequency lesion details, and redistributing image brightness. This effectively alleviates the issue of weak and difficult-to-recognize early lesion features. To enhance feature extraction with high efficiency and accuracy, the YOLOv11n backbone integrates an Adaptive Instance-aware Feature Interaction (AIFI) mechanism inspired by RT-DETR, achieving sparse and efficient feature interaction with minimal computational cost. Furthermore, an Adaptive Dual-Path Fusion Attention (ADFA) module is designed in the neck architecture to handle dense canopy occlusion. It performs multiscale convolutional feature extraction, cross-attention interaction, and adaptive dynamic fusion between feature pathways, thereby improving the discriminative capacity under occlusion and clustering conditions. RESULTS: Experimental results demonstrate that the CUG image enhancement framework significantly improves early lesion visibility and feature separability under complex illumination conditions. GDM-Net achieves accuracy and lightweight ability, demonstrating superior onboard deployment feasibility compared with other mainstream models. It achieves a 5.6% improvement in mAP@50 and a 6.5% improvement in mAP@50:95, while requiring only 7.5 GFLOPs of computational cost, addressing a key bottleneck in UAV-based GDM detection. CONCLUSIONS: GDM-Net provides an efficient and practical solution for UAV-based crop disease monitoring by integrating image information enhancement and improved lightweight detection. Its balance of computational efficiency and detection precision highlights its potential for broader applications in precision agriculture and intelligent crop health management systems.

Plant phenotyping relevance

UAV画像からブドウ葉の病斑・べと病状態を検出する画像強調および軽量深層学習手法を開発し、精度と計算効率を評価しており、植物病害表現型の取得手法が中心である。

Code and data availability

The supplied blocks describe a UAV image dataset of grapevine downy mildew (744 training / 83 test cropped images, 8,964/954 annotated lesions) and the GDM-Net model, but contain no data availability statement, no public repository deposit, and no author code/model release URL. The only URL present is the Creative Comm

No evidence-backed public reproduction asset is currently recorded.