【作者】 王晓；

【导师】 张小华； 杜海峰；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2019， 硕士

【摘要】 近年来,由于成像技术的发展,高光谱解混技术在越来越多的领域得到更加广泛的应用。在高光谱解混技术出现之前,高光谱图像都是采用“硬分类”的方法进行研究,这种方法将混合像元强制划分为某一类,显然是不准确的,从而产生了高光谱图像的“软分类”,即高光谱图像解混。传统的高光谱解混算法都是通过数学解析的方法求解,近几年由于深度学习迅速发展,神经网络框架广泛应用于各个领域,并且都取得良好的结果。由于高光谱解混的本质是一个函数逼近问题,而神经网络能够无限逼近任意连续函数,所以本文的研究内容主要是将神经网络应用在高光谱解混算法中,主要内容如下:1.由于高光谱解混是一个逼近问题,而神经网络可以拟合任意函数,此外高光谱图像在空间上具有相邻像元具有相似丰度的特点,因此本文提出了基于空间光谱相似性的卷积神经网络(CNN)的高光谱解混算法。对于一个高光谱像元,该算法将对应的丰度作为标签,通过卷积神经网络对丰度进行逼近来优化整个网络,提高解混的性能。该方法首先通过生成随机丰度得到对应的待解混的高光谱数据。然后对高光谱图像进行预处理,通过计算相邻像元的光谱角距离以及相对于中心像元的位置来设置相邻像元的权重。将高光谱数据作为训练样本输入到卷积神经网络中,使用反向传播算法对网络进行训练,通过训练好的网络对高光谱数据进行解混。实验结果证明了该算法的性能好于SUnSAL、MLP以及卷积神经网络。2.针对训练样本具有随机性这一问题,提出了基于改进的生成对抗网络(GAN)的高光谱解混算法。首先通过SUnSAL方法,对高光谱数据进行解混,得到初始丰度;然后将随机噪声输入生成对抗网络(GAN)的生成器中生成伪数据,将伪数据和上一步得到的丰度输入到判别器中判断真假,训练整个生成对抗网络,最终得到与初始丰度相似的丰度。通过计算光谱角距离(SAD)与相关系数可知,与随机样本相比,GAN生成的样本与原始的样本相似程度更高。生成器生成的丰度与端元矩阵线性相乘得到生成的高光谱数据,将其作为训练数据。将训练数据输入到卷积神经网络中,使用反向传播算法对网络进行训练,输入高光谱数据,网络通过训练来逼近其对应的丰度,通过训练好的网络对高光谱数据进行解混。通过实验结果得出,该算法的性能优于SUnSAL、MLP以及卷积神经网络。3.针对高光谱数据通常有几百个波段,存在大量的数据冗余,进而会引起计算速度慢和数据存储占用大量通道的问题,提出了面向高光谱解混的基于rmse指标的波段选择方法。首先,按照最佳组合指标(OIF)对输入的高光谱图像按降序排列,然后计算相邻波段子集的误差,最后设定阈值根据rmse指标去除高光谱图像中的冗余波段,完成高光谱图像的波段选择。实验结果表明,本文提出的针对高光谱解混的波段选择方法,明显优于一般的波段选择方法,并且提高了解混算法的运算速度,是更适合高光谱解混的波段选择。更多还原

【Abstract】 In recent years,due to the development of imaging technology,hyperspectral unmixing technology has been widely used in more and more fields.Before the emergence of hyperspectral unmixing technology,hyperspectral images were all studied by "hard classification" method,which forced the mixed pixels into a certain category,which was obviously inaccurate,thus producing the "soft classification" of hyperspectral images,namely,the unmixing of hyperspectral images.In recent years,due to the rapid development of deep learning,neural network framework has been widely used in various fields,and good results have been achieved.Since the essence of hyperspectral unmixing is a function approximation problem,and neural network can infinitely approximate any continuous function,the research content of this paper is mainly to apply neural network in hyperspectral unmixing algorithm,and the main content is as follows:1.Since hyperspectral unmixing is an approximation problem,and neural network can fit any function,in addition,hyperspectral images have similar abundance in adjacent pixels in space,so this paper proposes hyperspectral unmixing algorithm based on spatial spectral similarity convolutional neural network(CNN).For a hyperspectral pixel,the algorithm takes the corresponding abundance as a label and approximates the abundance through convolutional neural network to optimize the whole network and improve the performance of mixing.In this method,the corresponding hyperspectral data are obtained by generating random abundance.Then the hyperspectral image is preprocessed and the weight of adjacent pixels is set by calculating the spectral angular distance of adjacent pixels and the position relative to the center pixel.Hyperspectral data are input into the convolutional neural network as training samples,and the back propagation algorithm is used to train the network,through which the hyperspectral data are unmixed.Experimental results show that the performance of this algorithm is better than that of SUnSAL,MLP and convolutional neural network.2.Aiming at the randomness of training samples,a hyperspectral unmixing algorithm based on improved Generative Adversarial Network(GAN)is proposed.Firstly,the hyperspectral data are unmixed by SUnSAL method to obtain the initial abundance.Then,input random noise to generate pseudo-data in the Generative Adversarial Network(GAN),input the pseudo-data and the abundance obtained in the previous step into the discriminator to judge the true and false,train the whole counter network,and finally get the abundance similar to the initial abundance.By calculating spectral angular distance(SAD)and correlation coefficient,it can be known that,compared with random samples,GAN samples are more similar to the original samples.The abundance generated by the generator is linearly multiplied by the end element matrix to obtain the hyperspectral data,which is used as the training data.The training data is input into the convolutional neural network,and the back propagation algorithm is used to train the network,and hyperspectral data is input.The network approaches its corresponding abundance through training,and the hyperspectral data is unmixed through the trained network.Experimental results show that the performance of this algorithm is better than that of SUnSAL,MLP and convolutional neural network.3.Aiming at the problem that hyperspectral data usually has hundreds of bands and there is a large amount of data redundancy,which will cause slow calculation speed and occupy a large number of channels for data storage,a band selection method based on rmse index for hyperspectral unmixing is proposed.Firstly,input hyperspectral images are arranged in descending order according to the best combination index(OIF),and then the errors of subsets of adjacent bands are calculated.Finally,threshold value is set to remove redundant bands in hyperspectral images according to rmse index,so as to complete the band selection of hyperspectral images.The experimental results show that the band selection method for hyperspectral unmixing proposed in this paper is obviously superior to the general band selection method,and improves the operation speed of the mixing algorithm,which is more suitable for hyperspectral unmixing.更多还原