面向高光谱影像解译的无监督迁移深度表示模型与学习方法

It is crucial to efficiently use the massive hyperspectral images obtained by space-borne imaging sensors with costly expense. It is the key of improving the comprehensive applied ability of remote sensing and the challenge in the community of information science as well. This project aims at solving the curse of dimensionality and relatively small sample problems existing in feature representation, optimization and learning for massive, high-dimensional, non-structural and multi-model hyperspectral images. Inspired from the cognitive mechanisms of the human brain, which involving unsupervised learning and transfer learning, unsupervised transfer deep neural network (DNN) feature representation model is established for data reconstruction and data real distribution. An end-to-end unsupervised training style is designed and knowledge can be transferred between multi-source data sets effectively. Combing DNN with information theory and sparse representation theory, multi-objective models are built and efficient multi-objective optimization algorithms are designed for the proposed models. Feature dimension reduction and structure of networks can be optimized adaptively with a reasonable tradeoff. Based on the characteristics of hyperspectral images, some key learning problems in spatial-spectral feature representation for hyperspectral images are solved, which provides stable and robust approaches for spatial-spectral joint interpretation in multi-source complex hyperspectral scenes. It is expected to publish over 8 high quality papers, apply for over 4 patents, and jointly train 6 PH.D. or Master students.

花费巨大代价通过星载/机载传感器获取的海量高光谱遥感数据，如何有效解译与利用既是提高遥感综合应用能力的关键，也是信息科学领域的新挑战。本课题针对海量、高维、非结构、多模态高光谱遥感影像特征表示优化与学习中的维数灾难与相对小样本等问题，借鉴人脑认知的机理，从无监督学习与迁移学习出发，面向数据重构与数据真实分布建立无监督迁移深度神经网络特征表示模型，实现端到端的无监督训练模式以及多源数据集之间的知识迁移。将深度神经网络与信息论和稀疏表示理论结合建立多目标模型，设计高效的多目标优化方法，实现特征降维与网络结构的自适应权衡优化。结合高光谱遥感影像的成像特点与地物特性，求解高光谱遥感影像空谱联合特征表示中的关键学习问题，为多源复杂场景下的高光谱遥感影像的空谱联合解译提供稳健的方法。发表高水平论文8篇以上；申请国家发明专利4项以上；联合培养博士、硕士研究生6人。

本课题严格按照研究计划进行，主要研究成果包括：针对高光谱遥感数据标注训练样本匮乏的问题，构造全新的无监督深度神经网络模型，使其训练过程摆脱对标签样本的严重依赖，实现在分类等后续解析任务上在不降低过多精度的前提下，大大减少对标注样本的需求量；针对高光谱遥感数据的多模态特点，以及标注数据的匮乏问题，利用迁移学习原理，实现对已有数据集深度表达模型的知识迁移，即充分利用已有数据集的标注信息，获得在无标注信息或只有少量标注信息的全新数据集上的高效特征表达学习与推理能力；针对空谱遥感数据的非结构化特点，提出了基于注意力模型的空谱联合特征提取模型，能够通过自适应光谱和空间注意力机制降低特征提取过程中不相关光谱和空间信息的影响，得到高质量的空谱联合特征；同时还将所提出的特征提取方法应用到了变化检测，分类等任务中，获得了良好的性能。本项目共发表论文10篇，均发表在《IEEE Transactions on Cybernetics》，《IEEE Transactions on Geoscience and Remote Sensing》，《Pattern Recognition》，《Knowledge-Based Systems》等一区权威期刊上，申请国家发明专利2项，授权国家发明专利3项，联合培养博士2人硕士4人，研究成果获2022年陕西省电子学会自然科学二等奖（第一完成人），圆满完成了预期目标。

内容获取失败，请点击重试