Research On Hyperspectral Images Unmixing Algorithm Based On Optimized Nonnegative Matrix Factorization

Hyperspectral remote sensing images have data of tens to hundreds of narrow and continuous spectral bands,and contain rich geometric spatial information and spectral information.However,due to the complex land cover and low spatial resolution of the hyperspectral imaging system,the spectrum observed at each pixel is usually a mixture of several pure spectra,which can have a significant impact on the classification of hyperspectral images and ground object recognition.Hyperspectral unmixing is a key step in hyperspectral image analysis.The evolutionary algorithm is based on the population,iteration and random search mechanism,which is used to find the best solution to the related problems.Therefore,it has a good prospect to apply it to the mixing of hyperspectral images.Non-negative matrix decomposition aims at decomposing from multivariate data into partialbased representations,which is consistent with the objective function of hyperspectral decomposing.Therefore,the regularized non-negative matrix decomposition method for hyperspectral decomposing problems has been widely studied.On the premise of the robust non-negative matrix decomposition model of hyperspectral image based on evolutionary computing optimization algorithm,this paper improves the spatial resolution of hyperspectral image by blind separation algorithm.The main research work of this topic is as follows:(1)First of all,this paper improves the Bernstein search differential evolution algorithm,and proposes the Bernstein search differential evolution algorithm based on the structural gene strategy and the optimal guidance strategy.Bernstein search structure parameters of differential evolution algorithm is random,the cross process using Bernstein polynomial to control,so the Bernstein search differential evolution algorithm without cross process parameters,like other evolutionary algorithms,Bernstein,the success of the search algorithm to solve the problem is very sensitive to public parameter values.Therefore,in order to make the algorithm of pattern vectors can have better performance in the process of search,this paper puts forward the equation in the system of test mode vector to add the structure genetic strategy and the optimal intervention strategies,so as to improve the convergence speed of the algorithm as well as help algorithm to jump out the local extremum and improve the convergence precision of the algorithm.(2)Secondly,a hyperspectral image unmixing algorithm based on structural gene strategy and optimal guidance strategy based on Bernstein search differential evolution optimization algorithm is proposed.In this paper,Bernstein search optimization algorithm is used to replace the multiplication and iteration strategy in the robust non-negative matrix decomposition algorithm under the non-negative matrix decomposition model,which weakens the non-convexity of the objective function of the algorithm,enhances the global search capability of the robust non-negative matrix algorithm,and improves the spatial resolution of hyperspectral images.Using the structural gene strategy and the optimal guidance strategy of Bernstein search algorithm can not only guarantee the global search capability of solving the objective function based on the robust non-negative matrix decomposition algorithm,but also effectively avoid falling into local minima.Finally,the simulation data experiment and real remote sensing data experiment are designed,and the experimental results verify that the proposed algorithm has good unmixing performance.The experimental results show that the two optimization strategies can effectively improve the numerical function analysis and optimization ability of Bernstein’s search differential evolution algorithm.