Study On Compressive Sensing Of Hyperspectral Imagery Based On Linear Mixing Models

With the constant improvement of the spatial resolution and spectral resolution,the hyperspectral remote sensing data collected by imaging spectrometer have increased dramatically.The huge amount of hyperspectral data brings much pressures to data storage,transmission and processing of airborne or spaceborne remote sensing imaging system.Traditional hyperspectral remote sensing imaging is facing difficult problems to break through,such as high rate of sampling,massive data storage and transmission,etc.As a kind of novel sampling theory,Compressed Sensing(CS)subtly combined the data compression and the sampling process,realized the data acquisition with the sampling rate lower than the traditional Nyquist rate and accuracy reconstruction of small amount of observation data,reduce the requirements for the sensor and effectively avoided the software and hardware cost caused by pursuit of high resolution.The emergence of compressed sensing theory provides an effective way to solve the bottleneck problem met by the traditional hyperspectral remote sensing.It has become a research focus in hyperspectral data acquisition that the compressed sensing technology is applied to hyperspectral remote sensing imaging.The dessertation in-depth studies the compressive sampling mode and the imagery reconstruction algorithm of hyperspectral imagery based on linear mixing models,looking for the high efficient compressive sampling method and high precision and rapid compressed sensing reconstruction algorithm for hyperspectral imagery.Furthermore,the implementation schemes of compressed sensing imaging for the pushbroom and whiskbroom patterns widely used in hyperspectral data acquisition are also studed.The main work and innovations of this dessertation are shown as follows:1.A spatial-spectral compressive sampling scheme of hyperspectral imagery is proposed to replace the spacial or spectral compressive sampling for their simple and ineffective,based on analyzing the specialty of data collected by spacial compressive sampling.during the sampling stage,which compressive sample hyperspectral data firstly along the spatial dimension,then compressive sample spatial compressed data along the spectral dimension.Experimental results show that the spatial-spectral serial compressive sensing scheme is helpful for improving acquisition efficiency and reconstruction quality.2.Linear mixing models is used to reconstruct hyperspectral imagery from compressed data collected by spectral compressive sampling in order to avoid the huge size directly reconstruction of the original imagery.Reconstruction algorithms of compressive sensing based on unknown endmember,known endmember and linear spectral library mixing models are proposed.Hyperspectral imagery is separated into two small subset,abundance fractions and endmember signatures,by linear mixing models.The algorithm of hyperspectral unmixing is employed to estimate abundance,and the algorithm of sparse optimization is used to extract endmember,and then the hyperspectral imagery is reconstructed by the estimated abundance and extracted endmember.Experimental results show that,due to the small size of two estimated subset,the run speed of proposed reconstruction algorithm is greatly improved,and the algorithm can get a higher peak signal to noise ratio and better reconstruction effect.3.The distributed compressed sensing based on bands and pixels are proposed to solve the low accuracy and high computational complexity of hyperspectral imagery reconstruction.The hyperspectral image is divided into the key band images and the compressed sensing band images or the key pixel and the compressed sensing pixel during sampling,and different sampling methods are used for different types of data.The linear mixing models is used to separate different types of observed data during reconstructing,and then extract endmember and estimate abundance.linear spectral unmixing algorithm is applied to thansform the underdetermined optimization problem to the solving of over determined equation,which generate a substantial increase in the running speed and accuracy.According to the bands based distributed compressive sampling,a distributed compressed sensing reconstruction algorithm based on iterated prodiction is proposed to further improve the reconstruction accuracy.The experimental results of multiple hyperspectral datasets show that the reconstruction average signal to noise ratio of the two proposed schemes of distributed compressed sensing is much higher than the compressive projection principal component analysis algorithm and 3D compressive sampling algorithm,and the reconstruction speed has been increased by an order of magnitude than 3D compressive sampling algorithm.4.A spatial-pectral joint compressive sensing for hyperspectral imagery is proposed by constructing spatial-pectral joint compressive measurement matrix to make it easier to be realized by optical system.Through the analysis of the abundance characteristics for spatial compressed sampling data,a spatial measurement matrix with special structure is constructed for easy to realize by optical system.The original images are recoved by endmembers and abundances estimated by employing the linear mixing models during the reconstruction stage.The experimental results show that hyperspectral imagery colleted by spatial-spectral jiont compressive sampling and reconstructed separately by linear mixing models can achieve rapid and high precision goal for data acquisition and reconstruction.5.For the pushbroom and whiskbroom data acquisition patterns of hyperspectral remote sensing,the implementation scheme of compressed sensing imaging is studied.The optical system model of spectral compressed sampling for pushbroom hyperspectral imaging,pixel-based distributed compressed sampling for whiskbroom hyperspectral imaging and spatial-spectral parallel compressive sampling for pushbroom hyperspectral imaging are designed.The system models collect data by the photoelectric devices such as prisms,cylindrical lens,DMD,etc.The amount of data is sharply reduced without need the other compression processing.The designed system models save the storage space and transmission resources,and reduce the power consumption of the the data compression algorithm,and are advantageous to realize hyperspectral imaging for the airborne or spaceborne remote sensing platform.The simulation results show that the data colleted by designed imaging system models,although the reconstruction performance is slightly decreased,but still can restore the original data with high accuracy.