AWAVM-UNet: adaptive-weighted attention VM-UNet with multiscale attention feature aggregation for grape disease detection from UAV imagery

Grape diseases cause substantial economic losses worldwide, making accurate detection critical for effective control. UAV imagery offers a promising solution for automated disease surveillance, but detecting grape diseases from UAV images remains challenging due to high lesion variability, complex backgrounds (e.g., soil, shadows, overlapping canopy), and scale variation caused by changing flight altitudes. To address these challenges, we propose AWAVM-UNet (Adaptive-Weighted Attention VM-UNet), which integrates feature aggregation and channel-spatial attention. The model has three key components: (1) an AWA module that redesigns skip connections to fuse multi-scale encoder features using spatial and channel attention; (2) an MSCSA module that performs local multi-scale feature extraction via parallel group convolutions to handle altitude-induced scale variation; and (3) a CSA module with a learnable attention matrix that adaptively weights encoder-decoder features while suppressing background clutter. Experiments on a UAV-captured grape disease dataset show that AWAVM-UNet achieves state-of-the-art performance: 88.12% OA, 86.76% DSC, 76.58% IoU, and 88.39% Precision, outperforming CNN-based, Transformer-based, and Mamba-based methods. Ablation studies confirm the positive contribution of each component, and qualitative results demonstrate cleaner boundaries and fewer false positives, especially for small lesions at higher altitudes and in complex backgrounds. The proposed method provides an effective foundation for automated UAV-based vineyard disease monitoring.