Nonlinear Hyperspectral Unmixing Based On Neural Network And Differential Search Algorithm

Due to the limitations of hyperspectral image gradient descent unmixing algorithm,this thesis put forward two hyperspectral image unmixing algorithms.The first was nonlinear unmixing algorithm based on neural network and the second was hyperspectral image nonlinear unmixing algorithm based on neural network and differential search algorithm(DSA).We construct the objective function turning nonlinear unmixing problem into seeking best solutions to an optimization problem.By introducing the DSA,proposed algorithm transformed the parameters to be solved in the unmixing problem into the position parameters in the search process.And the mapping mechanism was introduced to meet the requirements of the abundance nonnegative constraint and sum-to-one constraint.The experimental results on synthetic and real scenes showed promising performances of the proposed frameworks compared with gradient unmixing algorithms.Therefore,in this thesis,the main work is as follows(1)First algorithm generated training samples to train the neural network using the generalized bilinear hybrid model.Use the trained neural network to estimate the nonlinear coefficient and abundance of the hyperspectral image each pixel,then reconstruct each pixel completing the nonlinear of hyperspectral image unmixing.(2)Second algorithm used the neural network to estimate nonlinear order.In nonlinear objective function optimization of the hyperspectral image unmixing algorithm,the gradient algorithm is susceptible to initialization and iterative step effect and convergence to local maximal value.By using the difference search algorithm which global convergence performance is better for optimization,we proposed the nonlinear hyperspectral image unmixing algorithm based on the neural network and DSA to overcome the nonlinear unmixing of the hyperspectral images.(3)According to the abundance nonnegative constraint and sum-to-one constraint of hyperspectral image pixel,non-negative constraint and abundance sum-to-one constraint were introduced into hyperspectral image unmixing algorithm to construct a new objective function.We put forward hyperspectral nonlinear image unmixing algorithm based on the neural network and difference search algorithm,and the difference search algorithm optimizes the objective function to achieve efficient nonlinear unmixing of hyperspectral image.