| 基于无人机影像与Mask R-CNN的单木树冠检测与分割 |
| |
| 作者姓名： | 黄昕晰 夏凯 冯海林 杨垠晖 杜晓晨 |
| |
| 作者单位： | 浙江农林大学信息工程学院 |
| |
| 基金项目： | 国家自然科学基金(U1809208);浙江省自然科学基金委员会-青山湖科技城管委会联合基金项目(LQY18C160002);浙江省科技重点研发计划项目(2018C02013);浙江省自然科学基金(LQ20F020005)。 |
| |
| 摘 要： | 对无人机遥感影像中单木树冠进行检测与分割并获取树冠冠幅与树冠面积参数,可以为城市中不同场景下的林业资源调查提供高效快捷的途径。以银杏树为研究对象,创建基于无人机遥感影像的银杏单木树冠数据集,并使用卷积神经网络Mask R-CNN算法结合正射影像图对城市中不同场景下的树冠进行检测和树冠边界勾绘以获取相关树冠参数。结果表明,加入无人机银杏树冠影像数据集训练后的网络模型,可以较好地适用于城市不同场景下的银杏单木树冠检测与分割。在4个测试场景下的86棵银杏单木树冠目标总体查准率达到93.90%,召回率达到89.53%,F1-score为91.66%,平均精度均值为90.86%,且可以提取到较为准确的银杏单木树冠的冠幅值与树冠面积,预测冠幅的平均相对误差与均方根误差分别为7.50%和0.55,预测树冠面积的平均相对误差与均方根误差分别为11.15%和2.48。将无人机影像与深度学习算法结合应用到城市林业资源调查中,可以得到较为准确的树冠检测与轮廓分割结果,有效地提高城市林业资源调查效率。
|
| 关 键 词： | 无人机 Mask R-CNN 树冠检测 树冠分割 冠幅 树冠面积 |
Research on individual tree crown detection and segmentation using UAV imaging and Mask R-CNN |
| |
| Authors: | HUANG Xinxi XIA Kai FENG Hailin YANG Yinhui DU Xiaochen |
| |
| Institution: | (College of Information Engineering,Zhejiang Agriculture and Forestry University,Zhejiang Provincial Key Laboratory of Forestry Intelligent Monitoring and Information Technology,Key Laboratory of Forestry Perception Technology and Intelligent Equipment,State Forestry Administration,Hangzhou 311300,China) |
| |
| Abstract: | Detecting and segmenting the single tree crown using the unmanned aerial vehicle(UAV) remote sensing images and obtaining parameters such as tree crown width(CW) and tree crown area(CA) can provide an efficient and fast way for forestry resources investigation in different urban scenes.At the same time,it can generate reference value for the estimation of urban tree health and growth status.The Ginkgo biloba tree in the city was selected as the research object,in which,the single G.biloba tree crown data set was obtained from UAV remote sensing images.The convolutional neural network Mask R-CNN algorithm and digital orthophoto map were used to detect the tree crown and draw the tree crown boundary to obtain the relevant tree crown parameters of different scenes in the city.The experimental results showed that the Mask R-CNN network model trained with the UAV G.biloba tree crown image data set can be well applied to the detection and segmentation of single G.biloba tree crown in different scenes in the city.In the four test scenes,the overall precision rate of the 86 single G.biloba tree crown targets reached 93.90%,and the recall rate reached 89.53%,Fl-score was 91.66%,and mean average precision was 90.86%.In addition,the tree crown width and tree crown area of the single G.biloba tree crown can be accurately extracted.The average relative error(ARE) and root mean square error(RMSE) were 7.50% and 0.55 for the predicted tree crown width,and 11.15% and 2.48 for the predicted tree crown area,respectively.It was shown that the application of UAV images and appropriate deep learning algorithm in the urban forestry resource investigation can obtain accurate tree crown detection and contour segmentation,and effectively improve the efficiency of urban forestry resource investigation.The method of this study can also provide technical support for the extraction of relevant tree parameters in the city. |
| |
| Keywords: | unmanned aerial vehicle(UAV) Mask R-CNN tree crown detection tree crown segmentation crown width crown area |
| 本文献已被 维普 等数据库收录！ |
|
