UAV-Based Intelligent Detection of Individual Trees in Moso Bamboo Forests With Complex Canopy Structure
Lujin Lv, Yinyin Zhao, Xuejian Li, Jiacong Yu, Meixuan Song, Lei Huang, Fangjie Mao, Huaqiang Du
Abstract
Detection of individual trees in Moso bamboo forests is critical to forestry resource management. However, accurate and rapid detection remains a significant challenge due to the high density of Moso bamboo forests and complex canopy structure. This study proposed a new canopy detection and counting method in Moso bamboo forests based on multiband images. First, this study used the dynamic thresholding method to extract the Moso bamboo forests' unique canopy hook tip features and coupled the original unmanned aerial vehicle (UAV) visible light images to construct multiband images. Then, this study utilized three object detection networks (faster R-CNN, YOLOv5, and YOLOv7) to detect the individual trees in Moso bamboo forests and to count the number of trees in the sample plots using the NMS method. This study assessed the new method's accuracy and compared the original UAV visible light images with multiband images in 84 Moso bamboo forest plots. The results showed that detecting the Moso bamboo forest canopy using multiband images improved accuracy in all three networks. On the test dataset, the YOLOv7 network using the multi-band images had the highest AP (89.15%) and R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> (93.17%), respectively, which were 3.18% and 15.5% higher than when using the original UAV visible light images. Faster R-CNN and YOLOv5 using multiband images also improved R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> by 7.3% and 7.2%, respectively. In addition, in the test dataset, YOLOv7 had the largest RMSE reduction of the three networks after using the multiband images, with a 37.93% reduction.