Hyperspectral Unmixing Using Autoencoder

Hyperspectral remote sensing images make important contributions to the research on the ground surface because of their rich spectral information.However,they also have the disad-vantage that cannot be ignored.The spatial resolution is very low,which results in multiple substances contained in a pixel.Therefore,decomposing various components in pixels is crucial for the use of hyperspectral remote sensing images.This is also the subject of this article,namely hyperspectral remote sensing images unmixing.For convenience,this paper refers to the hyperspectral remote sensing image as HSI in the following description.In past studies,a lot of works have been proposed to solve this problem.In this paper,we divided them into two categories according to whether they are based on the neural network or not,and name non-neural network methods traditional methods.Based on the existing work,this paper further explored HSI unmixing based on neural network,and proposed methods of HSI unmixing based on auto-encoders.The main works and conclusions are summarized as follows:(1)In this paper,a method based on partial auto-encoder is proposed to unmix HSIs.This method draws the inspiration from the traditional idea of non-negative matrix factorization so that link the linear minxing model of HSI with the structure of auto-encoders.The feature of samples is extracted by the encoder and is inverted by the decoder.The reconstructed data is obtained,the value of the loss function is calculated,and then is transmitted backward.The weight of the auto-encoder is updated by minimizing the loss function to find the suit-able weights of the auto-encoder.The physical meaning of the optimal decoder weights is the end members in this HSI obtained under this model.According to the mixing model,the input of the decoder is the abundance corresponding to each sample.This method can obtain the end members and abundance of the original image simultaneously,without the assistance of traditional methods.Experimental results show that the proposed method is superior to the traditional and the existing neural network methods.(2)A blind unmixing method of HSI based on sparse partial auto-encoder is proposed.In practical applications,the exact numbers of end members is usually few which are far less than the number of pixels,so the end members are sparse to the pixels.Moreover,not each pixel contains all end members,so the abundance of each pixel is also sparse.The auto-encoder structure is redesigned to this characteristic,and this design can determine the number of end members and express the abundance of each pixel more accurately.In this method,the number of end members is unknown,so the number of neurons in the hidden layer needs to be manually adjusted,and only enough neurons needed.The features pre-sented are sparse.This method with the ability of determining the number of end members is called blind unmixing.