【摘要】 【目的】针对在松枯死树监测实践中，从无人机航拍RGB影像中自动识别松枯死树漏检率高的问题，提出了一种生产应用场景下基于多色彩空间的YOLOv5松枯死树高精度自动识别新方法。【方法】利用无人机采集大面积松材线虫病发生林分的RGB图像，用Pix4Dmapper软件拼接，用Label Img开源软件建立VOC格式的松枯死树数据集，分别用Faster R-CNN、YOLOv3、YOLOv4、YOLOv5、SSD和Efficient Det等6种基于深度学习的目标检测算法对数据集进行训练和测试，以精确率、召回率、平均准确率以及F₁分数作为评价指标筛选出最优目标检测算法;然后将采集的RGB图像转换成LAB和HSV色彩空间图像，再将这3个色彩空间的图像分别用最优目标检测算法进行训练，得到目标在每个色彩空间的边界框，使用非极大值抑制算法对这些边界框进行处理，得到最优边界框实现松枯死树自动识别。【结果】6种算法均取得良好效果，其中YOLOv5模型为最优算法，其精准率、平均查准率和F₁分数在6种算法中均最高，分别达到97.58%、82.40%和0.85。通过3个色彩空间融合后，反映漏检情况的召回率由74.54%提高到98.99%，平均准确率提升至98.39%。【结论】基于多色彩空间的YOLOv5模型能够显著提高从无人机航拍RGB影像中检测松枯死树的精度，为松枯死树监测提供了有力工具，也有助于松材线虫病的防治。更多还原

【Abstract】 【Aim】 To address the issue of the high misdetection rate from RGB images taken by the UAV in the monitoring practice of pine dead trees, a new YOLOv5 model based on multi-color space is proposed for recognizing pine dead trees precisely. 【Method】 UAVs were used to collect RGB images of a large-area forest with pine nematode disease. The images were spliced with Pix4Dmapper software, and a dataset detailing pine dead trees in VOC format was established using LabelImg open-source software. Six object detection algorithms based on deep learning, namely Faster R-CNN, YOLOv3, YOLOv4, YOLOv5, SSD, and EfficientDet, were used to train and test the dataset. The optimal object detection algorithm was filtered using precision, recall, average precision（AP）, and the F₁ score as evaluation metrics. Then, the collected RGB images were converted into LAB and HSV color space images, and the three color space images were trained with optimal object detection algorithms to obtain the bounding boxes of the target in each color space. These bounding boxes were processed using non-maximum suppression algorithms, and an optimal bounding box was obtained to realize the automatic recognition of dead trees. 【Result】 All six algorithms have achieved good results, of which the YOLOv5 model was the optimal algorithm. Its precision, AP, and F₁ score were the highest among the six algorithms, reaching 97.58%, 82.40%, and 0.85, respectively. After the fusion of the three color space images, the misdetection was reflected in the recall rate, which increased from 74.54% to 98.99%, and the AP, which increased to 98.39%. 【Conclusion】 The YOLOv5 model based on multi-color space can significantly improve the accuracy of detecting dead pine trees from RGB images taken by the UAV. The proposed model is a powerful tool for monitoring dead trees and assisting in controlling pine wood nematode disease.更多还原