Research On Multiple Kernel Learning Based Classification Using Multi/Hyper-spectral Imagery And Lidar Data

Multi/hyper-spectral images acquired by remote sensing sensors can providesufficient spectral information of land-covers; LiDAR data provide informationmainly related to elevation distribution in spatial domain. The good complementaryrelationship between the above two data sources implies huge potential to makejoint use of them when confronting classification problem. However, data fromdifferent sources cause challenges to technique of combining them. This paper startsfrom analysis of characteristics of spectral images and LiDAR data, and mainlystudies classification by combined use of multi/hyper-spectral images and LiDARdata in applications of multiple kernel learning. The goal is to make better use of thementioned data sources and obtain better classification performance.This paper is mainly based on multiple kernel learning theories and itsapplication in joint use of multi/hyper-spectral images and LiDAR data whichinclude three aspects in detail.First, the basic principle of multi/hyper-spectral and LiDAR sensors along withcharacteristic of their data is introduced in this paper. Then noise reduction,rasterization and co-registration are carried out for corresponding data to make thempossible to be used together. To extract more information from the two data source,a series features, which include spectral bands, NDVI, nDSM, Morphologicalprofiles, intensity information of LiDAR, DSM generated by second return ofLiDAR, are extracted by using corresponding methods.Second, after make a brief review of basic principle and applications oftraditional kernel based methods, a multiple kernel learning framework based onboth scale and feature levels is proposed to tackle those heterogeneous features fromdifferent sources when make a joint use of spectral images and LiDAR data. Byadopting different strategies to solve multiple kernel learning problems in the scaleand feature level respectively, a multiple kernel based classifier with improvedclassification capability is obtained. The corresponding experimental resultsdemonstrate the effectiveness of our method in making joint use of spectral imagesand LiDAR data.At last, regarding to the problem of lack of labeled training samples comparedto the increasing number of features provided by different data sources, this paperreformulate the weight construction method in graph to adapt it to the situation ofcombining classification, then the improved graph is integrated with multiple kernellearning. After establish the corresponding objective function and give a solutionway of it, a multiple kernel learning classifier with semi-supervised capability is obtained and it is named graph Laplacian multiple kernel learning. This methodimproves the performance of multiple kernel learning based classifier by extractinginformation buried in unlabeled samples.