Deep Learning Individual Tree Segmentation Method For Coupling Airborne LiDAR And Digital Orthophoto Map

As one of the important parts of the global ecosystem,forests play an irreplaceable role in regulating global greenhouse gas levels,carbon cycles and other ecological issues.The traditional forest resource survey is to estimate the overall forest structure parameters and distribution status by sampling and measuring the parameters of individual trees in the sample area,which is limited by the topographical conditions of the forest,and the human operation will also bring many unavoidable errors,reducing the accuracy of the survey results.UAV remote sensing is an emerging active remote sensing technology that can quickly and accurately acquire large-area,multi-scale ground forest resource information and can reduce a large amount of fieldwork time.As the basic unit of a forest,the accurate and efficient extraction of single trees from UAV remote sensing data plays an important role in forest planning and management.The current individual tree segmentation algorithm based on UAV remote sensing data involves complex parameter settings,low automation and low segmentation accuracy,and more over-segmentation and under-segmentation will occur when encountering a high-density forest environment.To address the above problems,this paper takes a forested area in Xinlicheng,Changchun City as the study area,and uses the LiDAR point cloud data obtained by UAV with LiDAR and HD camera and UAV HD images as the data source,and proposes a deep learning method based on the rich features of individual tree canopy linearity and texture in the Digital Orthophoto Map(DOM)and the 3D structure information in the LiDAR point cloud for individual tree segmentation.A deep learning method of coupling DOM and point cloud clustering(hereafter referred to as the coupling method)is proposed for individual tree segmentation and 3D structure parameter acquisition based on the rich features of individual tree canopy linearity and texture in the DOM and the forest 3D structure information in the LiDAR point cloud.The following research work was conducted in this paper.(1)To obtain ground point clouds and non-ground point clouds by data preprocessing such as denoising and filtering of LiDAR point clouds,and to generate a Digital Elevation Model(DEM)with 0.5m×0.5m resolution by interpolation of ground point clouds using irregular triangular mesh algorithm.Generating a Digital Surface Model(DSM)with the same resolution as the DEM using ground point clouds interpolated with non-ground point clouds.The Canopy Height Model(CHM)is obtained by the difference operation of the above two models,and the processed original point cloud is normalized by combining with the DEM to obtain the normalized point cloud.(2)The orthophoto image is obtained by data processing such as inward orientation and relative orientation of HD image data,and the DOM is used as the data to produce the training sample dataset for deep learning,and the sample dataset is used for deep learning training of Mask R-CNN model,SSD model and Retina Net model.Based on the above models for canopy detection in DOM,the results show that the Retina Net model has the best canopy detection results,and the Mask R-CNN model and SSD model have poor detection results in the study area of this paper,and the detection results of the Retina Net model are selected in this paper as the canopy mask to participate in the subsequent individual tree segmentation.(3)The normalized point cloud is clustered by the canopy mask using the regional local maximum algorithm,and the point with the largest elevation is used as the seed point,and the normalized point cloud is segmented by using the distance discriminant clustering algorithm using the seed point to obtain the number of trees in the forest and the 3D structure parameters of individual tree.In order to evaluate the individual tree segmentation accuracy of the coupled method,the segmentation results of the currently commonly used watershed algorithm and layer stacking algorithm in three kinds of high,medium and low density sample subsurface were compared and analyzed with the coupled method,where the watershed algorithm is individual tree segmentation by CHM and the layer stacking algorithm is individual tree segmentation by normalized point cloud.The following conclusions were obtained from the above study.(1)The coupling method performs individual tree segmentation through the point cloud,avoiding the problem of loss of three-dimensional information during interpolation and rasterization of the point cloud.(2)Mutual validation by LiDAR point cloud combined with DOM and detection using Retina Net model,which avoids the process of constantly swapping thresholds and improves the automation of the algorithm.(3)The coupled method solves the problem of difficult individual tree segmentation in high-density forests with an F-Measure of 94%,compared with the watershed algorithm and the layer stacking algorithm,the F-Measure is improved by 6%-29% and7%-20%,respectively,that is,the individual tree segmentation results of the coupled method improve the correct rate of individual tree segmentation while maintaining a high detection rate,which can meet the modern forestry investigation for individual tree It can meet the needs of modern forestry investigation for individual tree extraction and achieve accurate and efficient individual tree extraction.