Refined Point Cloud Data Classification Of Full Waveform Airborne LiDAR In Forested Area

The last two decades witnessed the rapid development of Light Detection And Ranging(LiDAR)technology and its increasing applications in a wide range of fields with the superior capability of 3D measurement,spatial modeling and parameter inversion.Forests have been challenges to high resolution remote sensing for the complex vertical structures and variable properties of trees.Though the LiDAR technology has been employed in forestry to derive digital terrain models,classify tree species and extract vegetation parameters,the fundamental works for LiDAR data classification are still being affected with several systematic technical issues.The recent generation of LiDAR systems are benefiting from the huge performance enhancement,while new problems of data processing models are introduced as well.The characterization of targets is mainly restricted in the geometry and spatial topology of LiDAR point cloud,due to the inherent limit of single spectral wavelength,and physical observables from full waveform data are far from full utilization.Moreover,a forested area oriented technical framework for LiDAR data processing and classification is still absent to separate the ground,vegetation and above ground non-vegetation echoes fast and precisely.Geometric classifiers cannot guarantee the precision in the layer of 0.5 m above the ground,interfering the terrain model derivation indirectly.Non-vegetation objects like sparse buildings and powerlines are randomly mixed in the LiDAR data of forests,leading to reduced accuracy of forest parameters inversion.Based on these backgrounds,the paper conducted the study as follows,aimed at the refined full waveform data classification of airborne LiDAR.1)The correction of the term “range” in the direct geo-referencing model of LiDAR is proposed and the mathematical model of range ambiguities is systematically established.A novel approach of a priori terrain prediction is implemented to resolve the range ambiguity,coping with range ambiguities of new high pulse-repetition-rate airborne laser scanners,which caused incorrect point cloud geolocation especially in circumstances with variable terrain relief and high flying altitudes.It is proven that global digital terrain models like the ASTER GDEM are sufficiently accurate to provide range predictions for airborne Li DAR.This approach is not limited by LiDAR hardware mechanisms and has comparable robustness and higher portability than commercial solutions.This work guarantees the availability of point cloud data and is hence a significant precondition of further data classification.2)Gaussian decomposition,relative radiometric calibration and absolute radiometric calibration are implemented on full waveform LiDAR data.By geophysical reinterpretation,the “waveform ellipsoid” concept is proposed by forming an entity,which is quantized by waveform parameters and backscattering features,to characterize the spatial shapes of point cloud.Aimed at point cloud classification,the Waveform Ratio Index(WRI)is introduced by analyzing the statistics of echo broadening and backscattering characteristics to augment the separability between different echo types.Waveform Augmented Parameters(WAP)are gradually established by uniting the raw full waveform parameters,relatively and absolutely calibrated measurements and WRI.WAPs remarkably enlarged the feature space of single band LiDAR data and brought new ideas for higher level of objects characterizing.3)Last returns in forested area are classified to vegetation and ground by Random Forest algorithm with five combinations of waveform augmented parameters.Data of different incident angles are used to model classifiers,which showed that WRI occupy the highest classification accuracy,close to the combination of all WAPs.The overall accuracy can reach over 97% in proper observation conditions and is comparable to traditional geometric ground filtering algorithms.The accuracy of relatively and absolutely calibrated features are similar but both lower than that of WRI.The raw waveform parameters showed the lowest accuracy.It is also found that the increasing of incident angles can enlarge the variance of pulse widths,which consequently reduce the accuracy of all combinations,whereas the overall accuracy of WRI maintained over 82% if incident angles are less than 20°.This study confirmed the effectiveness of WRI as waveform augmented parameters and the superiority for LiDAR data classification,and is an innovative preparation for further refined classification.4)Fusing waveform augmented parameters and geometric ground filters,micro-terrain classification,a refined classification of low vegetation and ground returns in forested areas is put forward and buildings and powerlines are fast labeled with multiple parameter combinations in forests,constructing the ultimate refined LiDAR data classification framework.An approximating ground surface is firstly generated from last returns by weighted interpolation in the classical repetitive robust interpolation algorithm.A terrain buffer layer is then constructed by properly elevating the initial surface up.Robust features for classification are selected from frequency statistics and applied to the refined classification on vegetation and ground returns.It is found that the Gaussian Mixed Model can fit the distribution of near ground returns and therefor classification probabilities can be derived.The interference of low vegetation returns on the digital terrain model is eliminated to its maximum extent.This is essentially beneficial to complement geometric filters that can not cope with the confusions in the near ground layer below 0.5 m.In addition,labelling the buildings and powerlines purified the vegetation point cloud dataset to enhance forest parameters derivation.In the meanwhile,LiDAR equation sets for multiple returns are established.It is proven that the multiple returns phenomenon is a major cause to the poor robustness of waveform augmented parameters,which concludes suggestions for intensive full waveform LiDAR data investigations.In general,this paper is aimed at refined LiDAR data classification,and proposed new approaches on data processing,signal analysis and application strategies from the perspective of geometric localization,radiative transfer model and data utilization.Technical architecture from the preprocessing to the object classification is implemented and finalized on the precise digital terrain model derivation and refined point cloud data classification in forested areas.The work of this paper introduced new concepts and ideas to the forefront of Li DAR technology and created a solid technical foundation for precise parameter inversions of forests and vegetation.