Abstract: In response to the management difficulty of largescale Macadamia nuts orchards,  a forest Macadamia nut growth monitoring method based on improved YOLOv4 was proposed. Image acquisition was carried out in a Macadamia plantation, where three common forms of Macadamia presence were recorded to produce a VOC dataset and used for model training. Data augmentation was applied to classes with fewer samples to equalize the distribution of training samples. Improved on the basis of  the original YOLOv4 method, DenseNet121 network was used to replace the original backbone network, and Focal loss was used to optimize the classification loss function of the detection model, which effectively improved the detection model accuracy and alleviated the problem of unbalanced detection accuracy between classes. The experimental results showed that the improved YOLOv4 method had the highest average precision (AP) for each Macadamia nut form, compared to the YOLOv4 and YOLOv3, the mean average precision (mAP) of the detection model reached 93.33% and the detection speed reached 28.7 FPS, which achieved the realtime and efficient acquisition of growth information of Macadamia nut, such as Macadamia nut drop and disease in orchards, and provided a basis for Macadamia nut growth monitoring.

Key words:  , Macadamia nuts, orchard monitoring, deep learning, improved YOLOv4, object detection

摘要： 针对大规模澳洲坚果种植园管理困难的问题，提出一种基于改进YOLOv4的林地澳洲坚果生长监测方法。在澳洲坚果种植基地中进行图像采集，记录3种常见的澳洲坚果存在形式，制作VOC数据集并用于模型训练。对样本数量较少的类别进行数据增强，使训练样本均衡分布。在原始YOLOv4方法的基础上进行改进，用DenseNet121网络替换原来的主干网络，并使用Focalloss优化检测模型的分类损失函数，有效提升检测模型精度，同时缓解类别间检测精度不平衡问题。试验结果表明，与YOLOv4、YOLOv3方法相比，所提改进YOLOv4方法对每种澳洲坚果形式的平均精度（AP）均为最高，检测模型的平均精度均值（mAP）达到93.33%，检测速度达到28.7 FPS，实现对林地澳洲坚果落果、病害等生长信息的实时、高效获取，为精确监测澳洲坚果生长状态提供依据。

关键词: 澳洲坚果, 果园监测, 深度学习, 改进YOLOv4, 目标检测