Research On Feature Selection And Precise Classification Technique From LIDAR Data

LIDAR(Light Detection and Ranging) can obtain three-dimensional information ofintensive sampling points for a large spatial extent quickly,actively and automatically. Thistechnology remedies a defect that the traditional photogrammetric techniques only obtain fewfeatures of land-cover. Since LIDAR emerged commercially in mid-1990, this technique hasbeen effectively used for generating digital terrain model(DTM),natural hazard assessment,exploring roads and power lines, biomass estimation and so on. One of recent urgentproblems for LIDAR is to take advantage of multi-source information such as spectrum,texture, elevation and intensity provided by LIDAR system to obtain the distribution ofland-cover quickly and efficiently.In this dissertation, we mainly present feature extraction, the importance of features,building classifier in the field of land-cover classification by LIDAR data.Especially we tryto find the reasons why confusions exist in the result of classification and how to eliminationconfusions efficiently.In the past, when using LIDAR data to classify land-cover, feature selection mainly relieson personal experience and preference, resulting failing to achieve the optimal classificationaccuracy. To solve the problem, firstly, we should extract features as much as possible.Secondly the importance of different feature is achieved by analyzing permutation importancemeasure of Out-of-Bag(OOB) data.At last, the original feature space is replaced by newfeature space composed by the more important features.Strengths and weaknesses are analyzed among popular classification technique such asSupport Vector Machine(SVM), Markov Random Field(MRF), Dempster-Shafer Theory(D-S)and Random forests(RF).The most appropriate classification scheme is obtained byexperimental verification.There are some defects in the original result of classification, such as low accuracy, not matching the real land-cover characteristics, not meeting the observing habits of humanbeings and so on.In order to solve those problems, the types and locations of confusions areanalyzed, the reasons of confusions are found by using margin definition. Depending on thereason why different confusion appears, optimization algorithms are build based on spaceconstraints between local target and neighborhood to achieve better classification result.