多/高光谱图像融合分类的结构化低秩学习方法研究

Due to the low spatial resolution, noisy effect, and mixed pixels, the classification accuracies for hyperspectral image are difficult to further improve. Multispectral image aided hyperspectral fusion and classification is another kind of method to improve the classification accuracies and general performance. This proposal focuses on the multi-task structured low rank learning theory and related algorithm of "guide-feedback" coupled fusion and classification system for multispectral and hyperspectral image. The main contents are as follows. Firstly, we utilize probabilistic graph model to establish a novel superpixel sparsity measure by effectively exploiting both spatial and spectral correlation and contextual information in multispectral image and hypespectral image, and propose a spatial-spectral structured low rank based regularization method. Secondly, in the "fusion-classification" coupled system, we propose a new task-driven structured tensor dictionary learning method according to the "minimum reconstruction error" and "strong discriminative" criterion. Thirdly, we establish a structured low rank representation model of coupled fusion and classification for multispectral and hyperspectral image. Finally, we develop a series of efficient and effective algorithms for this coupled system. The research of this proposal would not only establish the solid fundamentals for fusion based classification, target detection, interpretation, but also have theoretical significance and application perspectives for multispectral and hyperspectral image.

由于空间分辨率较低、噪声干扰、混合像元现象明显，高光谱图像分类精度很难进一步提升。多光谱图像与高光谱图像空间-光谱信息融合分类的方式是进一步提升分类精度和泛化性能的有效方法。项目主要研究联合多光谱图像和高光谱图像融合与分类这一“指导-反馈”双向耦合系统下的结构化低秩学习理论与算法。主要内容为：挖掘多/高光谱图像的空谱相关性特征，利用概率图模型建立基于“超像素”同质地物的稀疏性度量与空谱联合的结构化低秩表示方法；针对“融合-分类”双向耦合系统，根据“重建误差最小”和“高可分性”准则，提出多任务驱动的结构化低秩字典学习方法；建立多光谱与高光谱图像空谱信息融合与分类联合优化的结构化低秩表示模型；利用算子分裂、交替迭代和GPU（CUDA）架构，设计融合和分类的并行优化算法。项目成果将为多光谱与高光谱遥感融合分类、解译、地物识别等后续应用提供有力支撑。

项目以多光谱/高光谱图像的融合-分类耦合问题为研究对象，基于结构化稀疏低秩理论探索高光谱图像融合、分类两个问题，以及耦合问题之间的“指导-反馈”效果。经过研究，本项目提出了一系列基于稀疏低秩特征字典学习的高光谱图像分类、融合的方法，主要包括：（1）提出一套基于高光谱图像原始空间先验信息（3DTV、交叉TV，差分空间低秩性等）挖掘的正则化融合方法；（2）提出一套子空间低秩分解的高光谱图像融合恢复方法；（3）提出一套基于稀疏回归框架的高光谱图像解混模型（L1-L2正则化）；（4）提出一套基于稀疏低秩表示与空谱核方法结合的高光谱图像分类方法。以上研究表明：高光谱图像恢复方法对后续的分类具有较好的改善效果，同时，基于分类后的结果，能够进一步改进图像恢复的精度，两者之间具有较好的“指导-反馈”效果，验证了项目提出的耦合系统的有效性。项目资助发表论文共计22篇，其中SCI期刊论文12篇，EI期刊论文2篇，国际会议论文8篇。培养硕士研究生5名，其中2名已取得硕士学位，3名在读。项目投入经费共计22万元，支出21.4576万元，各项支出基本与预算相符，剩余经费0.5424万元，将继续用于项目研究的后续支出。