基于非负自编码器及非负矩阵分解的高光谱解混

发布时间：2018-04-14 06:42

本文选题：高光谱解混 + 自编码器　；参考：《河北师范大学》2017年硕士论文

【摘要】：随着遥感应用逐渐趋向于定量化和精确化,高光谱解混作为高光谱遥感影像处理的关键技术之一,日益引起国内外学者的广泛关注。高光谱解混是指利用高光谱图像将混合像元分解为几种基本类型的地物光谱向量(端元),并求得这些基本地物所占比例(丰度)的技术。高光谱解混性能的提高,不仅有利于高光谱应用的发展,如地物的分类和识别、图像的解译和可视化、图像的增强和压缩等,也对地质勘探、农业监测和军事侦查等具有重要意义。本文对高光谱解混的研究现状进行总结与分析,并提出了两种新的非线性解混算法:(1)基于NNSAE-BP的高光谱解混结合非负稀疏自编码器(NNSAE)在挖掘数据内在结构、提取特征方面的优势,提出了一种基于NNSAE-BP的非线性高光谱解混模型。该模型的实现步骤包括两个阶段:解混模型的监督式学习阶段和高光谱数据的非线性解混阶段。最终将高光谱解混的问题转化为基于NNSAE编码模块的高光谱特征提取、以及基于BP神经网络丰度预测模块的端元丰度预测。在基于NNSAE编码模块的高光谱特征提取模型学习中,针对编码模块表示层节点的优选,提出了一种基于Otsu的编码模块表示层节点自适应选取算法,有效降低了模型计算的复杂度;利用BP神经网络采用小批量梯度下降法实现回归分析,并在目标函数中引入L2权重衰减项,防止过拟合。通过在真实高光谱数据库上的实验证明,该方法有效地提高了解混性能。(2)基于SSC-rNMF的高光谱解混从盲源分离角度出发,以非负矩阵分解(NMF)理论作为研究工具,同时引入端元的L2正则约束和丰度的L1/2正则约束,提出了一种基于平滑和稀疏约束的鲁棒非负矩阵分解(SSC-rNMF)方法。基于SSC-rNMF的解混方法是非监督式的,所需的先验信息仅为端元个数;采用块坐标下降法以乘法迭代的方式,实现高光谱图像的低秩分解;并引入符合高光谱解混特性的正则化,减小矩阵分解的解空间。不同于NMF,SSC-rNMF算法通过对非负的异常项进行乘法迭代计算,使其描述非线性因素的影响,提高模型的泛化能力和对噪声的抗干扰能力。最后,通过实验证明了该方法的有效性。
[Abstract]:As the application of remote sensing tends to be quantitative and accurate, hyperspectral demixing, as one of the key techniques of hyperspectral remote sensing image processing, has attracted more and more attention from domestic and foreign scholars.Hyperspectral demultiplexing refers to the technique of decomposing the mixed pixels into several basic types of spectral vectors (endelements) and finding the proportion (abundance) of these basic objects.The improvement of hyperspectral decontamination is not only conducive to the development of hyperspectral applications, such as classification and recognition of ground objects, image interpretation and visualization, image enhancement and compression, but also geological exploration.Agricultural monitoring and military investigation are of great significance.In this paper, we summarize and analyze the research status of hyperspectral demultiplexing, and propose two new nonlinear unmixing algorithms: 1) Hyperspectral descrambling based on NNSAE-BP and NNSAE-based non-negative sparse self-encoder (NNSAE), which has the advantages of mining the internal structure of data and extracting features.A nonlinear hyperspectral demixing model based on NNSAE-BP is proposed.The implementation of the model consists of two stages: the supervised learning stage of the model and the nonlinear de-mixing stage of the hyperspectral data.Finally, the problem of hyperspectral unmixing is transformed into hyperspectral feature extraction based on NNSAE coding module and endmember abundance prediction based on BP neural network abundance prediction module.In the learning of hyperspectral feature extraction model based on NNSAE coding module, an adaptive selection algorithm of coding module representation layer nodes based on Otsu is proposed, which effectively reduces the complexity of model calculation.BP neural network is used to realize regression analysis by using small batch gradient descent method, and L2 weight attenuation term is introduced into the objective function to prevent over-fitting.Experiments on the real hyperspectral database show that this method can effectively improve the performance of the SSC-rNMF based hyperspectral descrambling from the blind source separation point of view, using the non-negative matrix decomposition (NMF) theory as a research tool.At the same time, we introduce L2 canonical constraint of endmember and L1 / 2 canonical constraint of abundance, and propose a robust nonnegative matrix factorization (SSC-rNMF) method based on smooth and sparse constraints.The de-mixing method based on SSC-rNMF is unsupervised, and the prior information required is only the number of endmembers. The block coordinate descent method is used to realize the low-rank decomposition of hyperspectral images by multiplicative iteration, and the regularization which accords with the characteristics of hyperspectral descrambling is introduced.The solution space of matrix decomposition is reduced.Different from the NMF SSC-rNMF algorithm, it can describe the influence of nonlinear factors and improve the generalization ability of the model and the ability to resist noise by multiplying and iterating the non-negative abnormal items.Finally, the effectiveness of the method is proved by experiments.
【学位授予单位】：河北师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】