Hyperspectral Image Compression Method Research Based On Joint MSE And Classification Distortion Measure

With development of human science and technology, remote sensing technology is applied more and more. Hyperspectral image can collect spectral information from visible light band to infrafed band which has very high spectral distinguishing rate. So it is a great progress of remote sensing technology because it can solve many problems which multispectral image cann't solve. But at the same time, its huge data stream brings big difficulty to data storage and transmission. Because of these problems, it is very necessary to compress hyperspectral image data.While remote sensing image is different from general visual image; it is not only used for human visual system but also used for some special application such as classification. Reconstructed image quality is usually considered——how to get less MSE and how to get more SNR under some compression rate in remote sensing image compression research. Any compressed hyperspectral image must be used for some application; because of its application, MSE minimum criterion can not be treated as the only measure of compression; classification distortion measure should also be considered. The final goal is researching a better hyperspectral image compression method in the condition of classification application.In this thesis, the data character of hyperspectral image is first researched, which is the basis why hyperspectral image can be compressed. Then the relativity both between spatial pixels and between spectral bands is analyzed. Also we construe the high data dimension and entropy of hyperspectral image. Then the common compression measure is summed up especially compression measure based on MSE and classification. It is the basis of the next contents.Wavelet analysis is a very popular technology nowadays. It can collect more energy and has good character of time-frequency analysis. SPIHT algorithm is a kind of algorithm combined with wavelet's advantages well. Now, the 3D-SPIHT transformation compression method is brought considered character of hyperspectral image. This method realizes progressive transmission, gets best reconstructed image acquirable at every time. It is proved that it can get good reconstructed image quality in the experiment.Because hyperspectral image compression used for classification application usually only considers how to make mean square error minimum and is not related with the classification application, the new compression method based on joint MSE and classification distortion measure is brought forward. This method is mainly focus on that this compression is used to classification application, so both MSE and classification accuracy are used to measure the performance of compression system based on classification application. This kind of algorithm considers not only the quality of reconstructed image but also saves the class information better. Compared with vector quantization or 3D-SPIHT transformation coding, this algorithm advances classification accuracy greatly with keeping SNR and PSNR, realizes a hyperspectral image compression system based on classification application.