AWANet: Improved YOLOv11’s Adaptive Weighted Attention Network for Haystack Detection in Airborne SAR Imagery
Jianmin Hu, Cheng Chen, Jinting Xie, Boyue Li, Aguan Hong, Xiang Yi, Huanjun Chen, Yan Shen, Zhenyu Yang, Xinwen Zhang · IEEE Access · 2026
With the advancement of agricultural technology, China’s annual grain production has significantly increased, placing growing demands on efficient grain transportation planning. In order to in assist establish a fast and reliable statistical method for estimating grain storage distribution, thereby supporting decision-making in grain logistics, transportation, and resource allocation, this paper proposes an AdaptiveWeighted Attention Network (AWANet) based on deep learning for the detection of haystacks in Synthetic Aperture Radar (SAR) imagery. AWANet is built upon the YOLOv11 framework and incorporates a novel Weighted Attention Module (WAM) to enhance feature extraction and object modeling capabilities in Airborne SAR imagery by integrating three attention mechanisms. The introduction of trainable attention weights allows for dynamic balancing of these mechanisms during backpropagation. In the neck of the network, upsampling operation is replaced with deconvolution, and an Adaptive Attention Module (AAM) is proposed to compute adaptive attention weights based on intermediate-layer feature maps to mitigates the loss of low-level information during deep-layer processing and significantly enhances the network’s feature representation capability. Additionally, we adopt the Focal-EIoU loss function, which prioritizes difficult training samples and leverages rich geometric information. Experimental evaluation demonstrates that AWANet achieves 57.07% mAP@50-95 while maintaining a lightweight architecture, outperforming common models such as SSD, RT-DETR, YOLOv8, and YOLOv11. The proposed model consistently surpasses all compared methods in mAP@50-95, with a notable improvement of 4.79% over SSD in particular. These results fully validate the effectiveness and superiority of the proposed approach.