Research On The Methods Of Objects Classification By Combining Airborne LiDAR Point Clouds And High Resolution Aerial Images

With the constant improvement of resolution of remote sensing, we identify detailed information of object in a smaller scale. And LiDAR survey technique makes it possible to acquire the three-dimensional spatial information with high precision and makes information expression of remote sensing data closer to the perception of a human eye effect in reality. In face of multi-source data, the human brain can quickly analyze and extract the target of interest, but the processing mode of computer is not obviously intelligent enough, so information extraction technology based on multiple-data fusion still have a great space for development.The essay investigated classification extraction of surface features based on the fusion of LiDAR data and remote sensing data with high spatial resolution. Taking the advantage of features and the one which LiDAR data shows in three-dimensional spatial information, and combining LiDAR data with remote sensing images with a rich spectrum and texture information of high-resolution, we can combine the advantages of both and recover their own shortages. First, the essay summarized the principle and the process of data processing in LiDAR system, analyzed the LiDAR point cloud filtering method, and put forward an improved filtering method. Then distinguished ground points from non ground points are shown. Next, a further feature extraction of LiDAR point cloud data and image-data is shown, and it would be used for the fine classification of LiDAR point cloud data. Finally, the classification of high-resolution remote sensing images aided by LiDAR data were studied. Overall, the main work and the contributions of this paper are as followed:(1) This paper summarized and analyzed the basic principle of airborne LiDAR system, gave an overview of the characteristics of LiDAR data, and described the process of data preprocessing of LiDAR. Stating the basic principle and main methods of LiDAR data and image registration in detail, it provided the basis and theoretical origin for the subsequent data processing and information extraction from the fusion between LiDAR data and high resolution remote sensing data.(2) According to the characteristics of the LiDAR point cloud data which are discrete and blindness, this paper researched the method of LiDAR point cloud filtering. First it analyzed the basic principle and the difficulties of the process of filtering in LiDAR points cloud data, summarized and analyzed the shortage of existing filtering methods in solving difficult area filtering, proposed the LiDAR point cloud data filtering method based on multi feature fusion. Secondly it analyzed the attributes of the LiDAR point cloud data itself, and combined with the spectral information of the remote sensing images which are used in filtering process, adopting the multi-scale virtual grid data structure to organize the point cloud data and then using part of the initial separation of multiple echo information non-ground points in a single echo contain ground point with the last echoes of the point set, determine initial ground points according to the elevation of the texture characteristics and strength characteristics of point cloud; then it use the mean elevation grid, point cloud dispersion characteristics and spectral information to filter the laser spot. In the process of filtering, by changing the grid size to repeat the iterative calculation for further refinement of the filtering results, it generated the final DEM. LiDAR point cloud data filtering method was based on multi feature fusion, and joined the multiple features as the filter judging condition, avoiding the miss classification and misclassification in the single condition, to improve the overall accuracy of the filter.(3) In this paper, we worked on the LiDAR point cloud subtle category recognition technology based on results of sieving wave in order to meet the needs of the extraction of subtle object types using LiDAR point cloud data. We present a multi-feature point cloud classification methods based on feature weighted SVM combined with abundant image spectral and texture information to help to classify the LiDAR point cloud data. The study extracted features of point cloud and images based on the expression of discrete point set, and the SVM machine learning was used in classification. This paper improved the conventional RBF-SVM classifier using feature weight, and presented the image LiDAR point cloud classification method based on feature weighted SVM. The experimental results showed that this method could improve the precision and efficiency of the LiDAR point cloud classification.(4) For remote sensing image with high resolution, the object-oriented analyzing method which regards the object as the elementary unit of extraction and analyzing images has a clear advantage of taking full use of high resolution image geometrical and structural information in comparison with conventional method of pixel-based image classification. We studied the high resolution image classification with LiDAR data using the method of object-oriented analysis, and on the segmentation before classification basis, we first worked on image fusion, nDSM, and NDVI for multi-scale segmentation and generating the image object with homogeneity, then we worked on the fuzzy classification based on principles using the object's feature information. We presented the optimal segmentation scale selection method based on two different categories of parameters which are the reciprocal of standard deviation of the image object and the absolute value of mean difference between objects and their adjacent objects. We also build a different scales object hierarchy network structure, and analyze the definition of membership function for fuzzy classification, the establishment of fuzzy rule sets, and the process of classification. The experimental results showed that in comparison with LiDAR point cloud classification, high resolution image object-oriented classification with LiDAR data could extract more object classes with higher accuracy.The full text around the related technical issues of the fusion applications of LiDAR data and high-resolution remote sensing images, and study deep in specific fusion methods form three aspects, such as, LiDAR point cloud data filtering based on multi feature fusion, SVM classification of LiDAR data combined image features, and object-oriented classification of high-resolution image combined LiDAR data. Experiments results show that the proposed method can effectively improve the accuracy of classification results. The results of this study help to promote in-depth development of the classification technology combining LiDAR data and image data to some extent, and provide a reference for the development and application of multi-source data fusion technology.