高光谱遥感影像联合字典学习与分类研究

项目名称： 高光谱遥感影像联合字典学习与分类研究

项目编号： No.41471275

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 天文学、地球科学

项目作者： 杜培军

作者单位： 南京大学

项目金额： 90万元

中文摘要： 作为稀疏表达的重要内容，联合字典学习旨在建立具有重建能力和判别能力的字典和同步学习分类器，以提升分类性能。本项目针对高光谱遥感信息机理和影像特点，构建基于联合字典学习的高光谱遥感影像分类框架，实现三个有效的分类算法。首先，提出主动半监督联合字典学习方法，利用主动学习为半监督学习选择信息量大、无偏的未标记样本，使字典学习过程更有效，获得更好的字典和分类器。其次，构建先验知识引导、结构化稀疏诱导的联合字典学习方法，充分挖掘先验知识和规则，克服稀疏诱导规则的盲目性。第三，提出类别依赖的联合亚字典学习方法，利用较少的已标记样本获得对类别足够好的表达力，提高计算效率，避免字典冗余现象。项目从样本有限性、模型判别性和算法效率三个角度全面提升联合字典学习用于高光谱影像分类的性能。研究成果将拓展基于联合字典学习的高光谱影像分类方法，促进稀疏表达在遥感影像处理中的应用，推进高光谱数据的行业应用。

中文关键词： 高光谱遥感；半监督分类；稀疏表示；监督分类；主动学习

英文摘要： Joint dictionary learning is an important research direction in sparse representation, which allows for jointly learning the reconstructive and discriminative dictionary and classifier. This research focuses on hyperspectral imaging mechanisms and hyperspectral image (HSI) characteristics and aims at proposing a classification scheme for hyperspectral image based on joint dictionary learning. To this end, three joint dictionary learning algorithms for HSI classification will be proposed. Firstly, active and semi-supervised learning will be combined with joint dictionary learning, which will take full advantage of the unlabeled samples to learn the dictionary by adopting semi-supervised learning (SSL). In this phase, active learning (AL) methods are adopted to choose the most informative and unbiased unlabeled samples in order for SSL to work well. Therefore, the active semi-supervised dictionary learning method can obtain more powerfull dictionary and classifier. Secondly, the prior knowledge will be introduced for learning structured sparsity-inducing dictionary, which allows for a sufficient excavation of prior knowledge and rules to overcome the sightlessly design of structure. Thirdly, a class dependent joint sub-dictionary learning algorithm will be proposed, which needs less labeled samples in order to obtain adequate representation power. Moreover, the propsed method will greatly speed up the computation and alleviate the redundant phenomenon in the learned dictionary. To sum up, the research will focus on promoting the performance of joint dictionary learning for HSI classification in terms of the avaibility of training samples, the discrimination of the learned models and the computational efficiency. The outcomes of this research will expand the research of HSI classification based on joint dictionary learning, and further promote the application of sparse representation in remote sensing imaging. Last but not least, the research also has significant perspective in promoting the industrial application of HSI data set.

英文关键词： hyperspectral remote sensing;semi-supervised classification;sparse representation;supervised classification;active learning