Spruce counting method based on UAV visible images

ZHU Xueyan;ZHANG Xinwei;GU Mengmeng;ZHAO Yandong;CHEN Fengjun;School of Technology, Beijing Forestry University;Beijing Laboratory of Urban and Rural Ecological Environment;Department of Horticultural Science, Texas A&M University;Key Laboratory of State Forestry and Grassland Administration on Forestry Equipment and Automation;

seedling counting;;spruce counting;;deep learning;;unmanned aerial vehicle;;dense object;;YOLOv3

Since the traditional manual seedling counting is repetitive, boring, and subjective, it is an urgent need to develop a rapid and accurate alternative. To achieve this goal, visible light images of spruce plants caught by unmanned aerial vehicle(UAV) as objects and used deep learning technology to develop the fast and accurate spruce counting method. The DJI Phantom 4 UAV was used to acquire spruce plants' images, and 558 images were selected according to the principle of diversity. The contrast coefficient and zoom factor of pictures were adjusted to simulate spruce plants with different lighting conditions and different sizes, and the number of images were expanded to 1 674. The pictures were divided into training set(1 169 images) and test set(505 images) using a ratio of 7∶3. It was found that the version 3(YOLOv3) had the characteristics of multi-scale prediction, which can quickly and accurately detect targets with large differences in size, and was suitable for counting spruce plants of different sizes. Based on this, the YOLOv3 spruce counting model was constructed, and the training weight attenuation, initial learning rate, and batch size were set to 0.000 5, 0.001, and 64 based on experience. Spruce feature information and spruce targets were extracted by the Darknet-53 feature extraction module and the multi-scale prediction module respectively, and the number of spruce plants detected was the counting result. The performance of the YOLOv3 model were tested, and the mean counting accuracy rate was 90.24%; the root mean square error was 45.82, the underestimated percent, overestimated percent and difference percent were 15.47%, 19.25%, and 34.72% separately; and the processing speed was 0.415 s per image. Compared with the methods of fully convolutional networks(FCN) segmentation and Hough circle detection, the mean counting accuracy of the YOLOv3 model was 2.49% higher, and the root mean square error, underestimated percent, overestimated percent and difference percent were reduced by 29.32, 6.7%, 5.7% and 12.4% respectively. The results showed that the YOLOv3 model was an effective application in spruce counting.