基于深度信念网络的高光谱遥感影像变化检测方法研究

项目名称： 基于深度信念网络的高光谱遥感影像变化检测方法研究

项目编号： No.41501451

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 天文学、地球科学

项目作者： 黄风华

作者单位： 阳光学院

项目金额： 20万元

中文摘要： 针对现有高光谱遥感影像变化检测方法性能方面的瓶颈，提出了基于深度学习的高光谱遥感影像变化检测解决方案：首先，提出一种基于张量的多时相高光谱影像时-空-谱特征一体化模型（TFS-Cube），以充分利用高光谱遥感影像的底层特征并优化其组织方式，并针对 TFS-Cube张量数据的复杂性，提出一种基于改进型MBI的Tucker张量数据分解与降维方法(IPT-MBI);其次，设计一种基于张量的受限波尔兹曼机（Tensor-RBM），通过多线性运算实现TFS-Cube数据的采集、降维与非监督学习，并构建包含多层Tensor-RBM的深度信念神经网络（TBR-DBN），以实现对大量非标记变化/非变化样本的逐层深入训练；最后，采用支持张量机（STM）替代传统BP网络，并结合少量标记样本进行监督学习，提高TBR-DBN全局参数优化的效率。TBR-DBN可有效提升高光谱影像变化检测的精度、效率和自动化水平。

中文关键词： 高光谱变化检测；深度信念网络；受限波尔兹曼机；张量分解；支持张量机

英文摘要： Aiming at the performance bottleneck of existing change detection methods of hyperspectral remote sensing images, the change detection methods for hyperspectral images based on the deep learning are proposed as follows: first, a tensor-based integration model for time-space-spectrum features of multitemporal hyperspectral images (TFS-Cube) is proposed, so as to make full use of the underlying features of hyperspectral remote sensing images and optimize their organization mode. Additionally, because of the complexity of TFS-Cube tensor data, the method of Tucker tensors data decomposition and dimensions reduction is also proposed based on improved MBI (IPT-MBI). Secondly, a tensor-based restricted Boltzmann machine (Tensor-RBM) is designed to realize acquisition, reduction and unsupervised learning of TFS-Cube data by multilinear algebra, and a deep belief network which consists of multi-layer Tensor-RBM (TBR-DBN) is designed to realize the training of a large number of non-labeled changed /non-changed samples; finally, the support tensor machine (STM) is used to replace the traditional BP network, and it trains a small amount of labeled samples for supervised learning to improve the efficiency of the TBR-DBN global parameters optimization. TBR-DBN can effectively improve the accuracy, efficiency and automation level of hyperspectral image change detection.

英文关键词： Hyperspectral Change Detection;Deep Belief Networks;Restricted Boltzmann Machine;Tensor Decomposition;Support Tensor Machine