Research On Data-driven And Model-driven Hyperspectral Unmixing Methods

With the continuous advancement of spectral imaging technology and the continuous improvement of the spectral resolution of hyperspectral images,remote sensing science has entered a new era.But at the same time,due to the limitation of lower spatial resolution,the problem of mixed pixels appears,which has a great impact on the accuracy of hyperspectral image classification.In order to achieve the goal of refined classification of spectral pixels,hyperspectral unmixing came into being.Unmixing realizes the soft classification of mixed pixels by determining the end members and abundance of mixed pixels.In recent years,neural networks have achieved amazing developments in deep learning in various fields by learning the characteristics of data,and the abundance is just the low-dimensional feature representation of hyperspectral pixels.The main research content of this paper is to use deep learning to realize unsupervised and supervised unmixing of hyperspectral.The specific work is as follows:1.Taking into account the influence of noise in the image,a joint de-noising and unmixing network framework based on self-encoding network is designed,and a data-driven unsupervised network unmixing method is realized.The network design implements a custom network layer in strict accordance with the characteristics of abundance,so that the output meets the ANC and ASC constraints of the abundance.In the unmixed loss function,a regular term such as sparseness is constructed to suppress the noise of the reconstructed mixed data,and a self-encoding network with noise reduction and demixing is obtained.The process of denoising and unmixing the network promotes each other.The framework estimates the unsupervised abundance and endmembers,and then uses the estimated abundance and endmembers to reconstruct the denoised mixed pixels.After adjusting the training parameters,the framework is used to re-solve the denoised mixed pixels mix.The performance of unmixing is improved when the noise is suppressed.The experiment shows good unmixing performance on the synthetic data set and real data set polluted by high intensity noise.2.What usually happens in real scenes is spectral nonlinear mixing,and the application of linear mixing models in complex scenes is greatly restricted.A generalized bilinear model unmixing method based on autoencoder network is proposed for nonlinear mixing scenes.This method refers to the generalized bilinear model and divides the decoder of the autoencoder network into two parts: a linear mixing part and a nonlinear mixing part.Among them,the nonlinear mixing is adjusted by a set of learnable parameters,and the deep information of the mixed data is mined through network learning so that the method performs well on both nonlinear and linear data.Experimental results verify the effectiveness of the network on synthetic and real data sets,and show better performance than several unsupervised unmixing algorithms.3.The above two data-driven unmixing networks have achieved good unmixing results,but they also have some shortcomings: network topology design is difficult;network hyperparameter adjustment is troublesome.Based on the above reasons,we use the idea of unfolding the network to propose a deep unfolding network unmixing model based on the traditional SUn SAL-TV unmixing algorithm.This method combines the advantages of data-driven and model-driven methods,and uses the traditional unmixing model to solve the problem.The design problem of network topology,the use of supervised learning method to learn the parameter configuration in the traditional model,and greatly reduces the amount of data required to train the network,it has performed well in both traditional unmixing algorithms and data-driven supervised unmixing methods Performance.