Sparsity-constrained NMF Algorithm With Evolution Strategy For Hyperspectral Unmixing

Nonnegative matrix factorization has attracted increasing attention in hyperspectral unmixing due to the advantage that it’s a kind of blind source separation method with strong applicability and good explainability.The L1/2 sparsity-constraint nonnegative matrix decomposition has become a representative method to improve the unmixing accuracy by using the sparsity priori,but the optimization technique based on simple multiplication update rules makes the demodulation result easily fall into local minima and lacks robustness in practical applications.Aiming at the above problems,a sparse nonnegative matrix factorization optimization algorithm based on the combination of evolutionary strategy and multiplication update is proposed.Based on the effective initial solution set,the algorithm uses a combination of multiplication update rules based on alternating optimization techniques and a sparse measure-based coefficient matrix selection strategy in each iteration,and global optimization techniques based on differential evolution algorithms to find the best solution.We also use the adaptive differential evolution algorithm(IMMSADE)to adaptively improve the parameter control of the DENLMF algorithm,and proposes an IDELNMF method with better unmixing results.Based on the s sparsity-constraint nonnegative matrix factorization technology,the algorithm is improved by introducing a swarm intelligence optimization algorithm to solve the related problems of hyperspectral unmixing.The main innovations of this article are as follows:(1)Based on the sparsity-constraint L1/2NMF,the differential evolution algorithm in the swarm intelligence optimization algorithm is integrated,and the ability of global optimization is improved based on the original local optimization,so that the algorithm’s unmixing effect and processing speed have improved.(2)On the basis of the above innovations,this paper combines clustering and least squares when setting the initial population,so that the original data is more effectively used and a reasonable and efficient initial solution set is generated.Before the DENLMF algorithm,the data has a certain superiority and scientificity,and is verified through synthetic and real hyperspectral data.(3)In view of the shortcomings of DENLMF with many algorithm parameters,this paper also proposes a unmixing method(IDELNMF)that combines an adaptive differential evolution algorithm with sparse nonnegative matrix factorization.The automatic selection of algorithm parameters improves the ease of use of the algorithm in practical applications.