基于深度迁移网络的多模高光谱图像分类方法研究

Hyperspectral image classification is crucial for remote sensing based earth observation tasks. Its performance directly impacts the application and development of remote sensing. However, due to the redundancy of hyperspectral images and the insufficient of training samples, current classification methods are damaged by the over-fitting and dimensionality disaster problems. Moreover, due to the feature drift and the task inconsistency of multimode images, existing methods cannot transfer to multimode data. To solve these problems, this proposal studies the above issues from information mining, utilization and transferring. Firstly, in order to mine effective information from redundant data using a small training set, this proposal proposes a dual-channel deep deconvolution network to learn and optimize the spectral and spatial information of the hyperspectral image. Secondly, in order to efficiently utilize the insufficient training samples, this proposal proposes a semisupervised classification method based on the metric learning to perform classification. Finally, in order to transfer the proposed method in multimode data, this proposal explores solutions based on deep transfer networks, and proposes a domain adaptive method based on parallel transfer networks and a zero-time learning method based on multi-level transfer networks to complete classification of multimodal images. This project will realize accurate, stable, and continuous multimode data classification, the achievements and results will enrich and promote the research of hyperspectral image analysis, and provide theory and technique support for the applications of remote sensing in many fields.

高光谱图像分类是遥感对地观测中的关键问题，其精度直接决定着遥感技术的应用与发展。但是，受数据冗余大且训练样本不足的影响，现有分类方法受到过拟合与维数灾难等问题的困扰；同时，受多模图像特征漂移与任务不一致的影响，现有方法难以实现跨数据迁移。针对此，本课题拟从信息的挖掘、利用、以及迁移三个角度展开研究。首先，针对如何在训练样本不足时从冗余的高光谱图像中挖掘有效信息的问题，提出双通道反卷积网络以学习并优化空谱特征；其次，针对如何高效利用有限训练样本的问题，提出基于度量学习的半监督分类方法以实现源图像分类；最后，针对如何实现分类方法的多模数据迁移问题，探索基于深度迁移网络的解决方案，提出基于并行迁移网络的领域自适应方法与基于分级迁移网络的零次学习方法以实现多模图像分类。本课题将实现精准、稳定、持续的多模数据分类，其研究成果将丰富和完善高光谱图像分析方法，为诸多领域的遥感应用提供理论与技术的支撑。

高光谱图像分类是遥感对地观测中的关键问题，其精度直接决定着遥感技术的应用与发展。但是，受数据冗余大且训练样本不足的影响，现有分类方法受到过拟合与维数灾难等问题的困扰；同时，受多模图像特征漂移与任务不一致的影响，现有方法难以实现跨数据迁移。针对此，以实现多模高光谱图像分类为目标，课题组提出了基于双阶段关系学习网络的小样本分类方法解决样本不足问题，基于任务引导学习网络的小样本高光谱图像分类方法解决样本数量不均衡问题，基于领域自适应的高光谱图像分类方法实现跨数据分类，基于生成对抗网络的高光谱分类方法实现无监督跨数据分类，基于参数预测网络的高光谱图像特征提取方法有效学习高光谱图像特征，知识引导型高光谱图像分类方法实现高效的高光谱分类。此外，将提出的方法在无线感知、SAR图像目标识别与分类上、以及变化监测上进行了交叉应用。成果以论文形式公开发表，累计论文17篇，其中，SCI论文13篇，IEEE汇刊11篇，申请国家发明专利及软著12项，已获批国家发明专利2项，软著5项，项目的成果远超预期成果。