Band Selection Method Of Hyperspectral Image Based On The Improved Salp Swarm Algorithm

Hyperspectral remote sensing images have narrow spectral bands spaced from tens to hundreds of dimensions,containing rich spectral information,and are widely used in the fields of ground object detection,precision agriculture,and so on.However,the high dimensional characteristics of hyperspectral data can increase the cost of image processing,and also bring high data redundancy issues,affecting the efficiency and accuracy of subsequent image classification.Therefore,it is necessary to reduce the dimensions of hyperspectral images and remove redundant information before performing classification operations.In order to reduce the dimensionality of hyperspectral images,this thesis proposes an unsupervised band selection method based on the Salp Swarm Algorithm(SSA).Firstly,SSA is improved to solve the problems of slow convergence speed and easy to fall into local optimization.Subsequently,the improved algorithm is applied to the hyperspectral band selection process.The main research contents are as follows:(1)A binary Salp Swarm Algorithm based on golden sine and restart mechanism,called GRBSSA is proposed.In order to make SSA suitable for band selection,a transfer function is used to convert it into a binary Salp Swarm Algorithm(BSSA),and BSSA is improved to improve its convergence speed and accuracy.GRBSSA first uses a good point set strategy to replace the original random method to initialize the population of salps,making individuals in the population more evenly distributed in the search space,to improve population diversity;Secondly,the golden sine algorithm is introduced into the update formula of the leader position of salps to update the leader position twice,reducing its search range in each iteration to accelerate the convergence speed of the algorithm;In addition,a restart mechanism is introduced in the position update of the follower of salps,which is determined by the variance of the population fitness value.When the variance is less than a certain threshold value,the reverse solution is obtained for the previous position of the follower salps,and the original follower update formula is substituted to achieve a restart to prevent the algorithm from falling into local optimization.Optimization experiments were conducted on eight benchmark test functions,and compared with five existing binary group optimization algorithms.The results show that GRBSSA has good development and exploration capabilities,fast convergence speed,and high optimization accuracy.(2)A band selection method for hyperspectral images based on GRBSSA is proposed.Firstly,considering the differences between global and local features of hyperspectral images,the spectral dimensions of hyperspectral images are divided into subspaces,resulting in high inter band correlation in the same subspace and low inter band correlation in different subspaces.Then,an objective function for band selection is established to measure the information content of the band subset using average information entropy,measure the overall redundancy of the band subset using the average correlation coefficient between the selected bands,and limit the number of selected bands.The proportion of the number of selected bands to the total number of bands is limited by setting parameters.Using GRBSSA to optimize the objective function,band selection is performed in each subspace to obtain the final band subset.Finally,support vector machines are used to classify the selected wavebands at the pixel level.Three universal hyperspectral datasets are used to test the method proposed in this thesis,and compared with seven other waveband selection methods.The experimental results show that the waveband selection method proposed in this thesis has higher classification accuracy,demonstrating the effectiveness of the method in achieving hyperspectral image dimensionality reduction.