Reconstruction Of Hyperspectral Images With Compressed Sensing Based On Linear Mixing Model And Affinity Propagation Clustering Algorithm

With a dramatic increase in spatial and spectral resolutions of hyperspectral images,the amount of data captured by spectral imaging devices has surged.This poses significant challenges to airborne or spaceborne remote sensing systems in terms of data storage,transmission,and processing.To address these challenges and improve the efficiency of hyperspectral image storage and transmission,enhancing the reconstruction performance of hyperspectral images under low sampling rates or equivalent sampling rate conditions is a significant area of research in the field of hyperspectral image compression and sensing.Compressed sensing technology has found wideranging applications in the field of compression and reconstruction of hyperspectral images due to its ability to sample and compress signals using measurement matrices,and accurately recover the original signal from a small number of measurement values.However,the computational complexity of the recovery algorithm based on the optimisation algorithm is high and the computation time is long.Some scholars have proposed a reconstruction algorithm that combines linear mixing model and compressed sensing technology to compress and reconstruct hyperspectral images to reduce computational complexity,enhance processing speed,and simultaneously achieve high reconstruction accuracy.However,this approach faces two main challenges: first one,the low accuracy of estimating the endmember and abundance matrices of the reference band;second one,the large memory footprint and difficulty of implementation on optical devices for random measurement matrices.This study performs the following work based on the linear mixing model-based hyperspectral compressed sensing reconstruction algoritm:(1)In order to address the first issue,a compressed sensing reconstruction algoritm based on affinity propagation clustering algorithm and linear mixing model was proposed according to the spectral correlation of hyperspectral images.Specifically,the affinity propagation clustering algorithm is used to group the spectral bands during the sampling stage,with the clustering centers as the reference bands and the other bands within the group as non-reference bands.During the reconstruction stage,the number of endmembers is first estimated from the reference band,followed by the use of endmember extraction algorithms to obtain the endmember matrix.The abundance matrix is then estimated,and finally,the endmember matrix and estimated abundance matrix are used for reconstruction of hyperspectral images.Experimental results demonstrate that the algoritm proposed in this study achieves higher performance in the reconstruction of hyperspectral images than the linear mixing model based distributed compressed sensing method.The average signal-to-noise ratio of all reconstructed band obtained by the algorithm proposed in this study is 0.8784 d B higher than that of the comparison algorithm when sampling rate is 0.4.(2)To address the second issue,a 0-1 chaos-Toeplitz matrix is proposed.Chaotic sequences are generated using the logistic-map system,and then the chaotic sequences are changed to chaosToeplitz sequences by non-linear transformation.Finally,Finally,the chaos-Toeplitz sequence is transformed into a 0-1 chaos-Toeplitz sequence with only 0 and 1 elements by means of a discriminant function,then,0-1 chaos-Toeplitz sequence is transformed into a 0-1 chaos-Toeplitz matrix.Experimental results show that the 0-1 chaos-Toeplitz matrix proposed in this study has better performance than a random Gaussian matrix at the same sampling rate.The average signalto-noise ratio of all reconstructed band sampling by the improved measurement matrix of this study is 0.9648 d B higher than that sampling by Gaussian matrix when sampling rate is 0.3.The experiments also found that the measurement matrix affects the estimating accuracy of the endmember matrix of non-reference bands,which in turn affects the reconstruction accuracy of non-reference bands.