Detection and Recognition of Poppy in UAV Imagery Based on Deep Neural Networks

This study addresses the issue of poppy control by designing an improved YOLO algorithm based on UAV imagery for detecting and recognizing poppy targets in drone-captured scenes, enhancing its applicability in real-world scenarios. Traditional object detection models applied to poppy detection face challenges such as insufficient feature fusion, suboptimal attention mechanisms, and high false positive and false negative rates in complex environments. Additionally, these models struggle to balance detection accuracy with model efficiency. To address these issues, this study proposes an improved detection method, incorporating a custom-designed DC-C2f module to enhance feature extraction, an SD-attention detection head to focus on key target areas, and an SPD-Conv convolutional structure to optimize feature fusion and representation while reducing model complexity. In experiments, a DJI Matrice 300 RTK drone was used to collect 6,000 training images and 2,000 test images from rural landscapes under various lighting conditions, including sunny and cloudy days. Experimental results demonstrate that the proposed method achieves superior performance compared to mainstream algorithms, with a recall of 64%, precision of 84.8%, and mean average precision (mAP) of 72.9%, while maintaining a lightweight model architecture. These results significantly reduce false positives and false negatives, validating the effectiveness, superiority, and practical application potential of the proposed approach in poppy detection.