Classification And Modeling Of Ancient Building Components Based On Hyperspectral LiDAR

Chinese ancient buildings are the architectural and cultural treasures of our nation,which have high historical and artistic value.However,most ancient buildings are affected by natural factors such as wind erosion and oxidation,and are in an unsafe state and in need of intervention protection urgently.Manual surveying and mapping of ancient building components is likely to cause ’secondary damage’.How to collect and extract information from existing ancient buildings safely and efficiently is important for ancient building protection.Hyperspectral Li DAR（HSL）is introduced to obtain space-spectral point cloud data for the modeling of ancient building.It is difficult to classify wood component species in the normal modeling,so this research explore wood species classification based on HSL point cloud data,including the adaptive threshold finding method and the random forest（RF）classification algorithm based on feature sorting.The point cloud obtained by the conventional single-wavelength 3D scanner does not contain spectral domain information,and it is difficult to achieve classification in the model.In this study,the point cloud registration method is improved to realize the registration of HSL and 3D scanner point cloud.The space-spectral model was establish based on 3D coordinate and 101-dimensional spectral information,and then classification of wood species is verified base on spectral information.The main work and conclusions of this research are as follows:（1）The principle of the HSL system is analyzed in detail,firstly.Then the stability of HSL spectral information and the anti-interference ability of system（SNR）are mainly tested,also the accuracy of full waveform acquisition and scanning are verified.The results show that the spatial coordinates and spectral information collected by the HSL are accurate,which provides data for the classification of wood species and spatialspectral modeling of ancient building components.（2）Based on analyzing the spectral characteristics of timber,an adaptive threshold finding method and an RF classifier based on spectral feature sorting are proposed for wood species classification.Taking the average reflectance of full-band and the absolute difference of average reflectance and single waveband as the evaluation criteria,the adaptive threshold finding method is used to classify the ancient building components.In order to reduce the redundant information in the whole spectrum and improve classification speed,feature importance sorting is performed according to Gini index（Gini index）of RF output,the bands with higher feature importance are selected.The experimental results show that the classification time of RF classifier based on feature sorting is less than RF classifier data input with full-band spectral information,reduced by 7.32 s,while the accuracy rate reached 98.4%.Two methods proved the feasibility of classifying the wood species of ancient building components by HSL spectral information.（3）In the laboratory environment,3D scanner and HSL were used to collect point cloud data of ancient building components.After point cloud registration preprocessing,a bidirectional Kd-Tree based principal component analysis-nearest neighbor iteration（PCA-ICP）method is proposed to achieve the registration of two type point clouds,and then spatial-spectral information modeling is established.In order to improve the speed and accuracy of cloud registration,a strategy combining coarse and fine registration is adopted.First,the principal component analysis（PCA）method is used to obtain the principal axis direction of HSL point cloud,and then the rotation matrix is calculated to complete the rough registration of point cloud.The ICP algorithm of bidirectional KdTree performs point cloud fine registration,so that the closest points are found in one-toone correspondence.Compared with conventional point cloud registration algorithm,such as the combination of normal distribution transformation and nearest neighbor point iteration（NDT-ICP）,the registration time of this algorithm is reduced by 2720 ms,and the root mean square error（RMSE）is reduced from 0.00301 to 0.00292.Taking point cloud of 3D scanner as datum point cloud,calculate Euclidean distance of the corresponding coordinates of two type point clouds to obtain the closest point,which provide the HSL spectral information to 3D scanner point cloud.Coordinates combine the surface texture of the ancient building components to implement spatial model.Ancient building components in spatial-spectral model are classified by RF classifier based on spectral feature sorting,which prove that the spectral information of spatialspectral model is accurate,and it provides a new form for the information collection and preservation of digital ancient building.Figure [33] table [10] reference [62]...