Authors

Abstract

Identifying dry-direct seeded rice seedlings provides valuable information for field management. To address the challenges of seedling detection in cold-region dry-direct seeded rice fields, this study proposes an enhanced YOLOv11n-DF model. Key innovations include: 1) integrating DSConv into the C3k2 module to optimize phenotypic feature extraction, and 2) employing the FASFF strategy to improve scale invariance in the convolutional head. Experimental results show that the improved model achieves an mAP50 of 96%, with high recall, precision, and a processing speed of 251.5 FPS, outperforming the original YOLOv11n by 5 percentage points in mAP50, and surpassing YOLOv7–YOLOv10 in detection accuracy. The proposed algorithm effectively addresses challenges such as seedling occlusion and non-uniform distribution, offering a robust solution for automated seedling monitoring in precision agriculture.

Abstract in Chinese

本研究为应对寒地旱直播稻田的幼苗检测难度大的问题，提出改进YOLOv11n-DF方法。该方法引入了两个关键创新：1）DSConv引入C3k2模块以优化水稻表型特征提取，2）利用FASFF策略优化卷积头增强模型特征尺度不变性。实验结果表明，改进后模型的mAP50、召回率（R）、检测精度（P）分别为87.12%、77.07%和84.11%，处理速度为49.5 FPS，mAP50较改进前YOLOv11n提高4.28个百分点，在检测精度方面超过了同类目标检测网络（YOLOv7~10）。该算法有效地处理了旱直播秧苗相互遮挡和非均匀分布挑战，为精准农业应用中的自动秧苗监测提供了新的解决方案。