Research On Nonlinear Unmixing Method Of Hyperspectral Image Based On Bionic Intelligent Optimization

Hyperspectral unmixing is an important technique for identifying material features in hyperspectral images and estimating the abundance of material.Because the measurement accuracy of the hyperspectral sensor is limited,the sensor can only receive the mixed spectra of different material reflections,and it is difficult to satisfy the requirements of recognizing the material.At the same time,there are many local optimal solutions for the hyperspectral unmixing problem,which is a very challenging research problem in the field of remote sensing image processing.This thesis mainly adopts biology intelligent optimization algorithms for spectral unmixing.Firstly,spectral unmixing based on the higher-order nonlinear mixing model can effectively interpret complex spectral mixing processes and improve the performance of unmixing accuracy.However,the unmixing algorithm usually adopts the gradient optimization method with the single objective function of reconstruction error,which is susceptible to the outliers and is easy to fall into local optimum.Therefore,the thesis proposed a multi-objective hyperspectral unmixing optimization model based on the multi-linear mixed model,which has two objective functions of construction error and spectral angle mapper,and then optimized by the differential search algorithm.The experiment results show that,the proposed method can further improve performance of reconstruction error results with a better spectral angle mapper.Compared with the traditional gradient-based unmixing algorithm,the proposed method not only has a higher unmixing accuracy in reconstruction error,but also has a performance result in spectral angle mapper.Furthermore,this thesis finds that establishing a nonlinear spectral mixture model is important for spectral unmixing in complex scenario.There exist a few high-order spectral mixing models for characterizing complex scenarios,however,they typically have complex structures and related unmixing algorithms are easily fall into local optimum that limit its performance.In this thesis,the thesis proposes an exponential coefficient polynomial mixing(ECPM)model which considers the effect of multiple spectrum reflections and assumes that can use a single parameter to characterize complex spectrum reflection scenarios with the existence of high-order interactions among endmembers.The experimental results demonstrate that the proposed ECPM model can achieve better performance for spectral unmixing in terms of spectrum reconstruction error and spectral angle mapper simultaneously,in comparison with the existing baseline models.