| Forests are important part of the global terrestrial ecosystem and essential for our existence and development.Forest inventory is an important way to understand the status and dynamic changes in the forest resources.An efficient,fast and accurate access to different forest investigation factors has been an important work of forest inventory.Presently,Because of the low equipment cost,less required field work and so on;a variety of ground photogrammetry technology has been widely studied and used in forest inventory.However,these techniques have not been sorted out and compared under the same forest conditions,to determine the best scheme of forest inventory using ground photogrammetry.The purpose of this study was to examine the application of various terrestrial photogrammetry technologies in forest inventory,and determine the optimal production based on logical experimental results.The general production process of ground Photogrammetry for forest inventory is as follows:(1)scanning the forest sample plot through a certain scanning path to obtain images with a certain degree of overlap;(2)restoring the three-dimensional pose of each image during photography;(3)extracting the stand and forest parameters in the sample plot based on the image with pose.Therefore,this paper studies the various stages of technology in the production process.Mainly it studies three scanning paths: spiral,sunflower and simulated route method;on the image posture when restoring photography,it mainly studies the Structure from Motion(Sf M),Visual Odometry(VO)and Visual Simultaneous Localization and Mapping(V-SLAM)three technologies;on the extraction of standing trees and stand parameter,it mainly studies the image-based dense point cloud extraction method and the geometry-based extraction method constructed in this study.In order to compare the data of each stage in the data production process under the same conditions: in the experimental stage of the sample plot,this paper uses three scanning paths to scan the same sample plot to obtain three groups of original images;in the image pose recovery,this paper uses three pose recovery methods to recover each image pose for the three groups of original images of the sample plot,so as to obtain nine groups of image data with poses;Finally,nine groups of data are processed by the method based on image dense point cloud and the geometry based system constructed in this paper respectively,to obtain 18 groups of standing trees and stand data of a sample plot.In this paper,related experiments were conducted on 18 circular plots with a radius of 8 meters.The experimental results show that:(1)In terms of scan-path method,the RMSE of diameter at breast height(DBH)and Height of crown base(HCB)extracted by all methods in the spiral path was 0.75-1.79 cm and 0.12-0.32 m respectively;the RMSE of DBH and HCB extracted by all methods in the sunflower path was 0.80-2.22 cm and 0.13-0.39 m respectively;the RMSE of DBH and HCB extracted by all methods in the simulated route was 1.49-2.98 cm and 0.32-0.56 m respectively.The RMSE of the x-axis and y-axis of the stand positions extracted by all the methods in the spiral,sunflower and simulated route was 0.042-0.314 m,0.043-0.268 m,0.065-0.484 m,0.076-0.221 m,0.099-0.564 m and 0.109-0.935 m,respectively.The results showed that the spiral path was better than the sunflower path,while the sunflower path was better than the simulated route path.(2)In terms of the pose estimation methods,the RMSE of DBH and HCB extracted by all methods of Sf M were 0.75-1.55 cm and 0.12-0.36 m respectively;the RMSE of DBH and HCB extracted by all methods of V-SLAM were 0.87-1.77 cm and 0.14-0.39 m respectively;the RMSE of DBH and HCB extracted by all methods of VO were 1.45-2.98 cm and 0.27-0.56 m respectively.In Sf M,V-SLAM and VO pose estimation methods,the RMSE of the stand positions extracted by all methods in the x-axis and y-axis were 0.042-0.135 m and 0.043-0.113 m,0.075-0.178 m and 0.068-0.181 m,0.293-0.564 m and 0.221-0.935 m,respectively.The results showed that Sf M was better than V-SLAM,which in turn was better than VO.(3)In this paper,a system of extracting forest plot parameters based on Windows platform were constructed,which were used to extract standing trees and stand parameters from the image data of given orientation by geometric method.The system constructed the transformation matrix,from the initial world coordinate system of the image to the coordinate system of the sample plot,by a special chessboard placed in the center of the sample plot and the assumption that the standing trees are vertical to,so as to give each image scale information and pose in the sample coordinate system.Then,an algorithm for estimating the information of stand position,DBH and HCB was constructed based on the search of the specific points in it,which can further estimate the stand parameters.In the extraction algorithm based on dense image point cloud,it was necessary to use the transformation matrix obtained by the above geometric method to transform the image position to the sample coordinate system before the point cloud was densified,and then perform subsequent processing.In terms of the extraction methods based on dense point cloud and geometry,the RMSE of DBH and HCB extracted by all methods based on point cloud were 1.04-2.98 cm and 0.13-0.52 m respectively;the RMSE of DBH and HCB extracted by all methods based on geometry were 0.75-2.20 cm and 0.12-0.56 m respectively.Based on the point cloud and geometric methods,the estimated RMSE values of the positions of the standing trees in the x-axis and yaxis were 0.042-0.564 m and 0.043-0.935 m,0.042-0.433 m and 0.043-0.405 m,respectively.The results showed that the dense point cloud based on image was similar to the results obtained by geometric method.In summary,comparing the same image under the same forest conditions with different processing methods,it can be concluded that:(1)among the scanning paths,the spiral path is better,(2)in the pose estimation,Sf M algorithm has the best accuracy,V-SLAM algorithm has a little worse result but faster speed,which is suitable for real-time estimation,(3)among the two algorithms for extracting tree and stand parameters,the two algorithms have their own advantages and disadvantages.The algorithm based on dense point cloud has high accuracy but poor robustness,and needs to determine the scale and plot coordinate system by other external conditions.The geometry-based extraction system constructed in this paper,has high robustness and included a module to restore the scale and sample coordinate system. |
PDF Full Text Request