Research On Hyperspectral Remote Sensing Image Unmixing Algorithm Based On Convolutional Neural Network

Hyperspectral image data includes the spectral information of dozens or even hundreds of bands.Combining this detailed spectral information and spatial image information is helpful to achieve accurate ground object recognition.Therefore,hyperspectral remote sensing has widespread application and broad prospects in agriculture,forestry,the military,and many other fields nowadays.However,due to the limitation of the sensor,hyperspectral remote sensing data has the problem of high spectral resolution and low spatial resolution.At the same time,due to the influence of atmospheric scattering and other conditions,there are generally mixed pixels in hyperspectral remote sensing images,which significantly hinders spectral data analysis and is not conducive to subsequent tasks such as hyperspectral image classification and micro-target detection.Therefore,the hyperspectral image unmixing algorithm is still an important research direction in remote sensing.In recent years,neural network algorithms have developed rapidly,which have advantages in processing large amounts of data and excellent ability to fit nonlinear problems.As a kind of neural network algorithm,autoencoder can not only fit nonlinear problems,but also has excellent ability of data reconstruction,which can reconstruct hyperspectral image data.Convolutional autoencoder adopts convolution operation on the basis of autoencoder,which makes network parameters less and computation less,and is more suitable for dealing with problems with large data volume such as hyperspectral images.Therefore,this paper deals with the problem of nonlinear unmixing hyperspectral remote sensing images using a convolutional autoencoder network.The main contents are as follows:This paper proposes a hyperspectral nonlinear deconvolution algorithm NU-p CAE based on posterior adjustment for convolution autoencoder.The algorithm uses convolution operation and uses fewer parameters than the fully-connected autoencoder.The NU-p CAE autoencoder network consists of four parts: two encoders,a decoder,and a nonlinear output layer.Linear Encoder(LE)extracts the feature of the Linear part of hyperspectral image data.The Nonlinear Encoder(NLE)is used to extract the features of the Nonlinear part of hyperspectral image data and estimate the nonlinear parameters of the model.The output layer of the network uses the estimation of nonlinear parameters obtained by NLE to make a posterior adjustment to the model and then combines with the linear data reconstructed by the decoder to get the final output of the network.The unmixing network propagates back through the reconstruction errors of input and output,updates the weights of the network,and fits the nonlinear spectral mixing model.The weight of the decoder is the end element matrix,and the output of the linear encoder is the abundance matrix.In order to verify the estimation effect of the proposed algorithm on nonlinear mixing parameter b of hyperspectral images,a simulation dataset DS1 was generated,and Gaussian white noise with different SNR was added to it.The results show that with the increase in SNR strength,the performance of the NU-p CAE algorithm decreases slightly,but it can still meet the requirements of unmixing.In addition,in order to further verify the performance of the proposed algorithm,the other four classical unmixing algorithms are selected in this paper and compared with NU-p CAE on simulation data sets and real remote sensing data sets,respectively.The results show that the proposed method can obtain a more accurate abundance map,achieving an excellent end element extraction effect.In order to improve the unmixing accuracy of the NU-p CAE unmixing network,this paper also uses swarm intelligence optimization algorithm to optimize the hyperparameters of the network.The swarm intelligence optimization algorithm is simple in operation,does not need gradient information,has vital flexibility,and has good local optimal escape ability.It can solve problems that are difficult to solve by traditional optimization methods and has good performance in many aspects.An improved grasshopper optimization algorithm(EMGOA)based on dynamic double elite learning and sinusoidal mutation is proposed by adding two improved strategies to the grasshopper optimization algorithm.First of all,the algorithm can improve the local development ability of the algorithm and accelerate the speed of convergence through adaptive dynamic learning from two elite individuals.Secondly,in each iteration process,the sine function is used to guide the mutation of the current global optimal position to avoid the algorithm falling into local optimal and improve the convergence accuracy of the algorithm.In order to investigate the performance of the proposed EMGOA algorithm,experiments are carried out on ten benchmark functions.Experimental results show that the optimization performance of the EMGOA algorithm is significantly better than that of the basic GOA algorithm.Finally,the EMGOA optimization algorithm is used to optimize the hyperparameter of NU-p CAE unmixing network.Experimental results show that the convergence speed of the unmixing network after parameter optimization is faster.