Super-resolution Of Hyperspectral Image Via Superpixel-based Sparse Representation

Nowadays,hyperspectral images(HSIs)are more and more widely used in geophysics exploration,agricultural remote sensing,ocean remote sensing and environmental monitoring due to its rich spectral information.However,it is generally hard to obtain a high-resolution image with hardware limitation,and this will deteriorate the performance of HSIs on tasks above.Thus how to enhance spatial resolution of HSI via software method becomes an important problem in image processing.C.ompared with HSI,multispectral image(MSI)of the same scene generally have higher spatial resolution while spectral resolution is relatively lower,obviously the HSI and MSI of the same scene are information complementary.Thus we set out to research HSI super-resolution problem via fusing the spectral information of HSI and the spatial context of MSI to reconstruct high resolution HSI in this thesis.In this thesis,the research status of HSI super-resolution algorithm home and abroad is introduced,a superpixel sparse model and an improved superpixel group sparse model are proposed for HSI super-resolution.The main contents of this thesis are as follows:1.To better utilize the rich spatial information of MSI,we propose a HSI super-resolution method based on superpixel sparse reconstruction model.Firstly,we train a spectral dictionary from observed low spatial resolution HSI via online dictionary learning(ODL),and the spectral dictionary can be linearly transformed from HSI spectral dictionary.Once the dictionaries are obtained,the MSI can be segmented into superpixels and every superpixel can be jointly decomposed on MSI spectral dictionary to get sparse coefficients which carry rich spatial information.Finally,the target high spatial resolution HSI can be reconstructed by combining the HSI dictionary and coefficient.2.To take full consideration of inner structure of observed HSI,a superpixel based group sparse reconstruction model is proposed for HSI superresolution.Firstly,the observed HSI is clustered into several categories via kmeans++ algorithm and every category can train a subdictionary via alternating direction multiplier method(ADMM),MSI spectral subdictionary can be transformed from HSI spectral subdictionary.The oversegmented superpixels then choose a subdictionary which matches it best,and the sparse coefficient can be obtained via decomposing the superpixel on the selected MSI subdictionary.Finally,we can reconstruct high resolution superpixels via multiplying selected HSI subdictionary and sparse coefficient above.3.A software system is designed for hyperspectral image super-resolution methods mentioned above.The proposed methods are tested by Indian Pine,Cuprite Mine and Pavia Center.The experiments show that superpixels oversegmentation and subdictionary training can effectively utilize the rich spatial information of MSI and inner structure of observed HSI,thus get an improved reconstruct performance for HSI superresolution.