Research On Lossless Compression For Hyperspectral Images Based On Adaptive Filter

Hyperspectral image is a continuous narrow-band spectral image obtained by the airborne or space-borne imaging spectrometer on the earth’s surface in the visible to infrared region of the electromagnetic spectrum.It can be represented as a 3-D data cube containing 2-D spatial information and 1-D spectral information.Due to containing rich spatial and spectral information,hyperspectral image has a strong classification and recognition ability,and is widely used in geological exploration,environmental monitoring,resource management,military reconnaissance,meteorological observation,agricultural production and other fields.With the continuous improvement of spectral and spatial resolutions,the hyperspectral image data increases rapidly.The huge amount of data not only brings great pressure to the storage and transmission,but also restricts the promotion and application of hyperspectral image.Therefore,efficient compression is an urgent and fundamental problem in hyperspectral remote sensing technology.As a signal processing method based on linear model,adaptive filtering is very suitable for the interband linear prediction of hyperspectral image,and has a great potential in lossless compression of hyperspectral image.In this paper,through the analysis of characteristics for hyperspectral images,and combining with the characteristics of adaptive filtering theory and algorithms,the lossless compression of hyperspectral images is studied.The main research work and achievements of the paper are as follows:(1)A lossless compression algorithm for hyperspectral images based on clustered recursive least squares with adaptive band selection and adaptive predictor selection(C-RLS-ABS-APS)is proposed.For the problem that predictive accuracy of the recursive least squares(RLS)is affected by the degree of interband correlation and the stationarity of the prediction process,a lossless compression algorithm for hyperspectral images based on adaptive band selection and adaptive predictor selection is proposed.Firstly,aiming at the low spectral correlation of some bands due to atmospheric absorption,an adaptive band selection strategy based on the maximum correlation coefficient is proposed to improve the correlation between the bands to be predicted and their reference bands.Then,to solve the problem of low spatial correlation due to low spatial resolution,an adaptive predictor selection strategy based on clustering is proposed to improve the stability of the prediction process.In addition,in order to further improve the stability of the prediction process,a double snake scanning model and a recursive estimation method of local mean in similar neighborhood are designed.Experimental results show that the compression performance of the C-RLS-ABS-APS is superior to the state-of-the-art methods.(2)A lossless compression algorithm for hyperspectral images based on recursive least squares and adaptive threshold back pixel search(RLS-ATBPS)is proposed.Based on the detailed analysis of the special statistical characteristics produced by the correction process and their influence on the predictive effect of RLS,a lossless compression algorithm for hyperspectral images based on recursive least squares and adaptive threshold back pixel search is proposed.Firstly,the RLS predictor is used to generate the reference value of the pixel to be predicted.Then,for the special statistical characteristics generated by the correction process,an adaptive threshold back pixel search algorithm is proposed to search the candidate predictive values.In order to obtain the best prediction effect,an optimal threshold estimation method based on recursive error mean with coefficients is designed.Finally the final prediction value is selected by using the optimized minimum distance principle.The experimental results show that the adaptive threshold estimation method with estimation coefficient can approach the optimal threshold well,on the basis of the high prediction accuracy of RLS,so as to obtain the best compression results.(3)A lossless compression algorithm for hyperspectral images based on cascaded RLS-LMS prediction,namely clustered recursive least squares and least mean squares(C-RLS-LMS),is proposed.In view of the characteristics of high prediction accuracy of RLS but complex,and low prediction accuracy of LMS but simple,the cascade prediction model is introduced and extended,a lossless compression algorithm of hyperspectral images based on cascaded RLS predictor and LMS predictor is proposed.On the basis of clustering preprocessing,the algorithm adopts a three-level cascade prediction model.Firstly,the local mean(LM)predictor is used to remove the spatial redundancy,then the low-order RLS predictor is used to make spectral linear prediction of LM prediction error,and finally the high-order LMS predictor is used to make spectral linear prediction of RLS prediction error.The experimental results show that the algorithm C-RLS-LMS makes full use of the advantages of fast convergence of the RLS predictor and simple calculation of the LMS predictor,therefore achieves a good compression effect with lower computational complexity.(4)A lossless compression algorithm for hyperspectral image using clustered differential pulse code modulation and local mean(C-DPCM-LM)is proposed,and be implemented in parallel.Aiming at the problem that the spatial correlation is not utilized in the C-DPCM algorithm,a lossless compression algorithm for hyperspectral images based on C-DPCM combining with spatial correlation is proposed.Aiming at the easy parallelization of the algorithm,the parallel algorithm is designed by using CPU/GPU heterogeneous parallel computing model.In the preprocessing stage,the k-means parallel algorithm is used for clustering in first,and then the double snake scanning mode is adopted for serialization.In the prediction stage,LM predictor is used to eliminate the spatial redundancy in each cluster,and then the parallel algorithm of the regression prediction is used to make the optimal linear prediction of LM prediction error.The experimental results show that LM prediction significantly improves the prediction effect of C-DPCM algorithm,and GPU parallel computing significantly reduces the running time of the algorithm,thus achieving the fast compression with a high compression ratio.