Hyperspectral Unmixing And Application System Based On Sparse Representation

Hyperspectral remote Image(HSI)is spectral imaging data which can be captured using imaging spectrometer.HSI has hundreds wave bands and high spectral resolution,and it has been widely applied in many research areas,such as object classification and identification.However due to the restrict of space resolution of imaging spectrometer and the complexity of natural materials,one pixel in HSI usually contains many materials.Hyperspectral Image unmixing is one technology to discover the endmember and corresponding abundance in one mixed pixel,and it can be divided into two ways: blind hyperspectral image unmixing and semiblind hyperspectral image unmixing.For blind hyperspectral unmixing technology it needs to extract endmember form a hyperspectral image first and then calculate corresponding abundance.Semi-blind hyperspectral unmixing approach has recently emerged as a new tool which depends on a prepared spectral library,and all its need is to calculate abundance map.The quality of spectral library plays an important role in semi-blind unmixing.however due to the existence of spectrum mismatch between the actual spectral signatures in the scene and which in spectral library,the result of semi-blind unmixing has received some effect.In this thesis,we focus on spectral library based semi-blind hyperspectral image unmixing.We study the classic collaborative sparse regression,and then propose two novel hyperspectral image unmixing algorithms by modeling the spectrum mismatch as sparse error vectors.The main contribution of the thesis are as follows:(1)The unmixing algorithm(CLSUnSAL)based on collaborative sparse regression is analyzed.Inspired by bounded error of spectrum mismatch which proposed by dictionaryadjusted nonconvex sparsity-encouraging regression(DANSER),this thesis proposes a novel unmixing algorithm based on dictionary sparse pruning and collaborative sparse regression(DSPCSR).This algorithm solves the spectrum mismatch problem effectively by modeling the spectrum mismatch error as a sparse vector.Both simulate and real experiment results show that proposed DSPCSR can achieve good performance.(2)Spatial correlation of pixels lying in the homogeneous regions of hyperspectral images is analyzed.Inspired by alternating direction sparse and low-rank unmixing algorithm(ADSpLRU),this thesis proposes a new unmixing algorithm: simultaneously dictionary sparse pruning and low rank for hyperspectral image unmixing(DSPLR).This algorithm takes into consideration the truth that pixels belong to the local area contain same materials,hence low rankness property is applied to abundance matrix.Experiment shows that DSPLR can achieve better unmixing precision.And the abundance maps achieved by DSPLR have good robustness and piecewise smoothness in local region.(3)A spectral library based semi-blind hyperspectral image sparse unmixing prototype system is developed.The system contains many kinds of hyperspectral spare unmixing algorithms,spectral library pruning algorithms and a complete hyperspectral unmixing procedure.The procedure includes: input hyperspectral image,prune spectral library,run unmixing algorithm,and output abundance map finally.The system may have widely application prospect.