| Hyperspectral image(HSI) is one kind of remote sensing image. However, it is quite different from other remote sensing images such as visible, infrared. That's because HSI have an extra dimension, namely spectral dimension, which contains a large amount of spectral information, moreover, the spectral resolution of HSI is strong. So it is widely applied in global observation, target detection etc. Although high spectral resolution of HSI bring tremendous benefit to the interpretation of remote sensing information, it also bring difficulty to transmission and storage of HSI. That's because all the huge advantages of HSI are exist at cost of its high amount of data. Therefore, it is significant to study on hyperspectral images compression and encoding.We just do our research based on such a background. In order to provide some theoretical support to hyperspectral image compression, we do some research on the characteristic of spectral correlation, spatial correlation and information entropy of hyperspectral images in the start of this paper. The analysis to HSI spatial characteristic show that each band of hyperspectral images is intraband correlation, however its correlation much lower than normal digital images, so it is not a proper way to expand the compression algorithm of normal 2D image to 3D hyperspectral images directly. The research on spectral correlation of HSI reveal that the spectral correlation between two bands of hyperspectral data is much stronger than the spatial correlation within a band, so it means that we must focus on the removal of spectral redundancy in the process of hyperspectral compression. The study on entropy of HSI indicate that the compression ratio of hyperspectral image could not be too high, because each band of HSI contains rich information, we will lost much key information if compression ratio is high.Secondly, in order to remove spectral correlation of hyperspectral images, we proposed a three-dimension spectral modeling based on hyperspectral image compression algorithm. By this algorithm, an initial prediction is implemented in the spectral dimension of hyperspectral data to obtain the initial prediction value, and then we establish a three-dimension spectral modeling according to the neighborhood of predicted pixel. The established model is used to modify the initial prediction value and obtain final prediction value. The final prediction value is more close to the true value of a pixel, so the accuracy of prediction is improved. There is still spatial redundancy after the process of spectral prediction, so we carry out wavelet transform to the residual images and encode them by 2D-SPIHT . We use reconstruct images signal to noise and entropy of residual images to assess the performance of compression method. Experimental results show that the proposed algorithm is outperform two classical algorithms–the optimal spectral DPCM and adaptive linear prediction algorithm.Third, we design a spectral-spatial hybrid predictor. Although 3D spectral model based compression algorithm is better than some classical methods, it is also computational complexity, moreover, the decorrelation of spatial and spectral needs two steps, so we proposed a spectral-spatial hybrid predictor on the basis of several spatial predictors and spectral predictors. Experimental results show that the accuracy of hybrid predictor is higher than spatial predictors and spectral predictorsThe ultimate goal of research on hyperspectral images compression is solve the problem of transmission and storage. So we construct a hyperspectral image compression hardware system in the last of our paper. This system decorrelate the spectral correlation by FPGA, then remove the spatial correlation and encoding residual images via ADV202. The output of this system is compression codestream. The decompression result by software in computer shows that the system compressed hyperspectral images efficiently, and the compression performance of hardware system is same as the simulation by software. |
PDF Full Text Request