黄丽明, 王懿祥, 徐琪, 刘青华. 采用YOLO算法和无人机影像的松材线虫病异常变色木识别[J]. 农业工程学报, 2021, 37(14): 197-203. DOI: 10.11975/j.issn.1002-6819.2021.14.022
    引用本文: 黄丽明, 王懿祥, 徐琪, 刘青华. 采用YOLO算法和无人机影像的松材线虫病异常变色木识别[J]. 农业工程学报, 2021, 37(14): 197-203. DOI: 10.11975/j.issn.1002-6819.2021.14.022
    Huang Liming, Wang Yixiang, Xu Qi, Liu qinghua. Recognition of abnormally discolored trees caused by pine wilt disease using YOLO algorithm and UAV images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(14): 197-203. DOI: 10.11975/j.issn.1002-6819.2021.14.022
    Citation: Huang Liming, Wang Yixiang, Xu Qi, Liu qinghua. Recognition of abnormally discolored trees caused by pine wilt disease using YOLO algorithm and UAV images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(14): 197-203. DOI: 10.11975/j.issn.1002-6819.2021.14.022

    采用YOLO算法和无人机影像的松材线虫病异常变色木识别

    Recognition of abnormally discolored trees caused by pine wilt disease using YOLO algorithm and UAV images

    • 摘要: 松材线虫病是一种传播速度快的毁灭性森林病害,利用无人机遥感及时对松材线虫病病害木进行监测,是控制松材线虫病蔓延的有效方式。该研究利用YOLO算法自动识别无人机遥感影像上的松材线虫病异常变色木,利用深度可分离卷积和倒残差结构改进YOLOv4算法,提高了识别的精度和效率。比较Faster R-CNN、EfficientDet、YOLOv4和YOLOv5与改进的YOLO算法的速度和精度,并分析了改进的YOLO算法在参与训练区域和未参与训练区域的异常变色木的识别效果。试验结果表明,改进后的YOLO算法的平均精度为80.85%,每个迭代周期的训练时间为164 s,参数大小为44.2 MB,单张影像的测试时间为17 ms,表现优于Faster R-CNN和YOLOv4,但与EfficientDet和YOLOv5相比有优有劣,综合比较这4个指标,改进算法在检测速度和检测精度上的表现更为平衡。未参与训练区域异常变色木的F1分数(84.18%)略低于参与训练区域(87.92%),但基本满足异常变色木的监测要求。相似地物、林分郁闭度、坡向和分辨率会对识别精度产生影响,但影响较小。因此,改进的YOLO算法精度高、效率高,可用于松材线虫病异常变色木的快速识别,并且对未参与训练区域异常变色木的识别具有较高的适用性。

       

      Abstract: Abstract: Pine wilt disease is a fast-spreading and destructive forest disease which can damage the entire pine forest in a short time. The key to control the disease is to identify abnormally discolored trees in a quick and accurate way. In this study, a fixed-wing Unmanned Aerial Vehicle (UAV)equipped with a professional true-color camera was used for image acquisition, and a deep learning algorithm-YOLO (You Only Look Once), was adopted to detect the images of abnormally discolored tree. The test was conducted on the platform using a RTX 2080 GPU with 8GB memory in the same parameters. The study area was divided into one training area and two testing areas, of which the training area and the testing area 1 were located in Xinjing Mountain, and the testing area 2 was located in Baota Mountain. In order to improve the efficiency and accuracy of the algorithm, depthwise separable convolution and inverted residual block were employed to improve YOLO algorithm, and the upward transmission of the low-level features of location of the Neck was removed. In order to verify the performance of the improved algorithm, the improved algorithm was compared with Faster R-CNN, EfficientDet, YOLOV4 and YOLOv5 in terms of accuracy precision, training time of per epoch, size of parameters and testing time of per image. Specifically, the accuracy precision of the improved YOLO algorithm increased to 80.85%, which was 1.25%, 3.02% and 3.49% higher than Faster R-CNN, YOLOv4 and YOLOv5 respectively,and was 0.42% lower than EfficientDet; the training time of per epoch of the improved algorithm was 164 s, which was 187, 172 and 115 s shorter than Faster R-CNN, EfficientDet and YOLOv4 respectively, and was 26 s longer than YOLOv5; the parameters size of the improved YOLO algorithm registered 44.23 MB, which was 477.35 and 199.69 MB smaller than Faster R-CNN and YOLOv4,and was 29.62 and 17.27 MB bigger than EfficientDet and YOLOv5; the testing time of per image of the improved YOLO algorithm decreased to 17 ms, which was 68, 33 and 7 ms less than Faster R-CNN, EfficientDet and YOLOV4 and was 9 ms longer than YOLOv5. Therefore, comprehensive evaluation of the four indexes showed that the improved YOLO algorithm performed better. In order to verify the applicability of the same model in different areas, the model built on images of training area of Xijing Mountain was employed to detect images in both testing areas of Xijing Mountain and Baota Mountain. The results showed that in Xijing Mountain, the improved YOLO algorithm correctly detected 582 trees, wrongly detected 109 trees and missed 51 trees among 633 abnormally discolored trees by pine wilt disease in total. The precision, recall and F1 reached 84.23%,91.94% and 87.92% separately. In Baota Mountain, the same algorithm successfully detected 708 trees, incorrectly detected 109 trees and missed 157 trees in a total of 865 object trees. The precision, recall and F1 were 86.66%,81.85% and 84.18% in separate. Although the F1 of Baota Mountain was slightly lower than that of Xijing Mountain, it basically met the detection requirements of abnormally discolored trees caused by pine wilt disease. The F1 of two testing areas showed that the model possessed certain adaptability and meant the same model can be used to identify target trees in different areas, which can reduce the modeling time and improve the efficiency of recognition. Through the results of the recognized images, it could be found that the similar ground features, canopy density, slope aspect and resolution ratio were the main factors that had an influence on the accuracy of detection. In automatic detection, the similar ground features would lead to acquire features similar to abnormally discolored trees and the canopy density, slope aspect and resolution ratio would acquire insufficient feature, which would easily lead to misjudgment and missed detection of abnormally discolored trees. Although these factors would affect the accuracy of detection, the impact is relatively small. In conclusion, the improved YOLO algorithm can automatically identify trees attracted by pine wilt disease in a more accurate and efficient manner, which was conducive to monitoring of pine wood disease rapidly, knowing the incidence timely and providing objective data for the prevention and control of pine wilt disease.

       

    /

    返回文章
    返回