Building Extraction And Scene Understanding For Hyperspectral Remote Sensing Imagery And High Spatial Resolution Remote Sensing Imagery

The building is one of the important indicators to measure the development of the city.Automatic batch extraction from remote sensing imageries will undoubtedly provide more rapid and efficient geographic information for the development of the city.Compared with traditional artificial methods,it is more time-saving and labor-saving.Scene understanding is a new hot spot with the accumulation of remote sensing data.It is used to solve how to make computers "understand" imageries as human brains,and automatically identify,classify,and manage massive imageries.Based on the understanding of the scene of a building,the method of automatically identifying,analyzing,and understanding imageries related to buildings is to be sought,providing scientific basis for city informationization,fine management,and smart city construction.The main technical methods and innovations of this article are as follows:Firstly,an object oriented multi-scale segmentation and low-level multi feature extraction and fusion method for the original hyperspectral high spatial resolution imagery is proposed.Aiming at the single low level features describe the object not fine situation,using the optimal segmentation parameters of original hyperspectral and high spatial resolution imagery multi-scale segmentation,and then extract the representative types of low-level features and high dimensional low-level feature fusion,each object.The high-dimensional feature better depicts the characteristics of the terrain.This process is to reduce the dimension of hyperspectral remote sensing imagery and reduce the amount of computation.However,the high resolution remote sensing imagery is a dimension enhancement process,which improves the expression accuracy of the feature.Secondly,a method of sparse representation based on high dimensional low-level features are proposed,and Fisher constraints are added to the sparse representation process.The building is the use of high dimensional low-level feature extraction on object level,combined with the sparse representation classification method,and joined the Fisher criterion in the dictionary learning process,solve the difficult problem of information capture relative differences in similar categories of traditional sparse representation method of the dictionary.The results of batch recognition and classification show that there are differences in the data sources,imaging quality and the complexity of the terrain,which leads to the diversity of the classification results.In general,the imagery recognition and classification results with good imaging quality and low terrain complexity are better than those of other imageries.But one to many sample processing,large area,multi form,high difference of building extraction in the previous research results belong to small number of cases.Thirdly,a middle level semantic scene classification method based on RBF-NN model is proposed.Neural network has powerful learning function.Combined with middle level semantics,it realizes hyperspectral and high spatial resolution sensing imageries in the same system,and analyzes the scene classification of multi-source remote sensing imageries.The results show that better classification results can be obtained by using fewer samples,and the effectiveness of the method is verified.It is also different from the traditional method to classify the scene by using the global or local statistical feature information of the whole imagery.Fourthly,initially proposed from the original high-spectral high spatial resolution imagery low-level multi-feature extraction and fusion,to the building extraction analysis,and finally to a more complete system based on the scene classification of the building.Combined with experiments,it verifies the practicability and effectiveness of the system,and provides a way for future feature extraction,classification and scene understanding of multi-source remote sensing imagery.