Research On Feature Extraction And Classification Methods Of Hyperspectral Imagery

With the progress of space technology and information processing technology,remote sensing technology developed rapidly in the last half of the 20th century.Combined with imaging and spectral detection,a new branch of remote sensing,which is called hyperspectral remote sensing,comes into being.Compared with other kinds of remote sensing images,there are advantages in hyperspectral images,such as narrow spectral range,continuous wave band,approximately continuous spectral information of objects and strong ability in object recognition.In recent years,hyperspectral remote sensing technology is widely used in fields of investigation and development of land resources,monitoring of natural disasters,and etc.With wavelet transform,signal information can be separated into a high frequency part and a low frequency part.Multi-fractal can provide a accurate description to image details in different scales.As details and differences among pixels mainly exist in the high frequency spectrum information and the correlation is in the low frequency part,a classification method of hyperspectral images based on multi-scale features and multi-fractal analysis is proposed in this dissertation.The main work of this thesis includes:The background knowledge of hyperspectral remote sensing image and its development both at home and abroad are introduced,which includes the current hyperspectral feature extraction method,hyperspectral pattern recognition method and the commonly used indicators to evaluate the efficiency of classification on hyperspectral images.1.The hyperspectral images are preprocessed,which includes image denoising and mixed pixel separation.The image denoising is completed by a discrete wavelet transform,and the separation of mixed pixels is based on multi-scale wavelet features,preparing for the subsequent pixel unmixing,classification and recognition.2.The fractal theory and the principle of multi-fractal spectrum are described in detail,and the multi-fractal property of hyperspectral is analyzed.Then multi-fractal spectrums of hyperspectral images are extracted.3.The endmember extraction and abundance estimation of hyperspectral images are completed by the separated pure pixel set and mixed pixel set,and pixel unmixing is realized.4.Experiments are made on AVIRIS database.Two groups of comparison tests are designed.One compares the features presented in this thesis with SPPI+PCA,and the other compares the classification efficiency using different classification algorithms as nearest neighbor algorithm and support vector machine algorithm with the same features proposed in this dissertation.According to the results,the proposed method in this thesis is more effective.