Research Of Improving The Accuracy Of Land-cover Fast Classification Method Based On LIDAR Data

Airborne Light Detection and ranging system(LIDAR) is becoming a key data source of remote sensing. Fast classification of LIDAR data is the base of applying LIDAR data into actual project, so it occupies an important position in the field of LIDAR data processing. But in the practical application, the information provided by LIDAR data can hardly response the feature of land-cover fully and accurately, and it is often difficult to establish discriminant rules in the case of lacking prior knowledge, which leads to the accuracy of classification results cannot meet the high precision needs of the application. This seriously restricts the development and application of LIDAR technology. In this paper, we aim to research on how to effectively improve the classification accuracy of the algorithm, while being sure that the computational time of the method is fast, the main contents of this paper are as follows:(1) Research on the features of LIDAR data and characteristics of land-cover objects. A brief introduction of the structure and components of LIDAR system was carried on, the principle of data acquisition and physical meaning of objects were analyzed deeply, and which kind of objects can be distinguished by each kind of LIDAR data were analyzed. The identification characteristics of five pre recognition classes were analyzed. Factors which affect the accuracy of classification result were analyzed from aspects including data acquisition, data transformation and classification methods, and the corresponding solving method were proposed.(2) Research on LIDAR data land-cover multi classification method based on fuzzy DSmT method. First the concepts and types of the basic DSmT were introduced, then the method of fuzzy DSmT was put forward according to the way of how to reduced factors affecting the classification accuracy. The uncertainty in the LIDAR data and conflict caused by fusing multi-sources data were quantified by constructing fuzzy sets and conflict sets. Multi features were fused by building three kinds of fuzzy probability distribution functions and the probability redistribution function. Finally a variety of classification results were gotten by making decisions, and the effectiveness of proposed method is analyzed through comparison.(3) Research on a single land-cover extraction method based on the fusion theory of probability distribution. Single land-cover extraction method only focuses on a particular object, and it is also another way to improve the classification accuracy. Firstly the related concepts and fusion theory of probability distribution is introduced. Take automobiles as an example for feature extraction, four features of vehicle were constructed for recognition, the ratio of length and width and proportion of region intensity were quantified by building probability distributions with probability-possibility transformation method based on the uncertainty of relationship between characteristic value and vehicle possibility. The distribution fusion rule was constructed based on T-mode, S-mode and average operator, and two kinds of distributions were fused. A more reliable result was obtained than that got from single distribution. Finally, the recognition result was gotten by making the final decision, accuracy rate, leakage identification rate and error identification rate were defined to evaluate the effectiveness of this method.