高光谱遥感图像亚像素-像素-超像素协同分类研究

Hyperspectral image classification is one of the most critical technologies in the field of earth observation by remote sensing. It provides practical value of high importance for various applications, e.g., precise agriculture, resource and environmental monitoring. However, due to the complex spectral-spatial properties and insufficient training samples, it is difficult to obtain high classification accuracies. To solve the above issues, this program plans to do research in the following aspects. First, effective subpixel-level, pixel-level and superpixel-level feature extraction algorithms will be designed, to better reflect complex spatial properties (e.g., shape, structure, texture) and spectral properties (e.g., spectral shape, spectral mixture), which will lead to the construction of multi-view images. Then, spectral-spatial adaptive sparse representation based joint multi-view representation method will be proposed, which integrates the self-correlation and discriminate information within multi-view images. By this way, multi-view images can be represented in a more compact and more complement way, thus, complex spectral-spatial properties can be more effectively characterized. Finally, multi-layer information interaction based collaborative classification method will be proposed. Here, multi-layer information will be jointly exploited, including the intrinsic correlation information between labeled and unlabeled samples, as well as the feature-layer and decision-layer complementary information in multi-view images. This will greatly contribute to the improvement of hyperspectral image classification performance, especially in the case of limited training samples. The execution of this program will open up new ideas for precise land-cover identification and hypersectral image classification, and promote broader applications of hyperspectral remote sensing.

高光谱遥感图像分类是遥感对地观测的核心技术之一，在精准农业、资源与环境监测等领域具有十分重要的应用价值。高光谱遥感图像空谱特性复杂，样本标定困难，导致图像分类精度低。针对上述难点，本项目拟开展以下研究: 1）挖掘亚像素级、像素级和超像素级空谱信息，研究能反映复杂空间(形状、结构、纹理等)和光谱(光谱形状、光谱混合等)特性的多级特征提取方法，构建亚像素-像素-超像素多视图；2)分析多视图内在相关性和表征差异性，提出空谱自适应稀疏表示的多视图联合表示方法，实现结构紧凑、信息互补的多视图表示，克服复杂空谱特性难以高效表示的问题；3)建立多视图内已标记和未标记样本结构关联模型，结合多视图在特征层和决策层的互补信息，提出多层信息交互的协同分类方法，突破标定样本不足导致的分类性能瓶颈，提升分类精度。本项目研究成果将为高光谱遥感图像的精细地物分类提供新思路，为高光谱遥感在多个领域的广泛应用提供技术支持。

该项目面向精准农业、资源与环境监测等领域实际应用需求，结合多视学习、主动学习、流形学习等先进机器学习理论，围绕高光谱遥感图像亚像素-像素-超像素协同分类理论与方法开展深入研究，重点在高光谱图像复杂空谱特征高效提取与联合表示、多视图空谱特征协同分类方面取得了重要研究进展。首先，结合高光谱图像中光谱信息、端元丰度信息以及空谱结构相似信息等先验知识，提出了亚像素-像素-超像素多层级空谱特征提取方法，建立了高光谱图像空谱结构保持的特征选择模型，构建了亚像素-像素-超像素多视图；其次，建立了联合局部视图辅助和超图自学习的判别性空谱特征表示模型，提出了基于超像素布朗算子的空谱特征联合表示方法，实现了多视图空谱特征的高效联合表示；最后，深入挖掘样本结构分布的全局与局部关联特性，提出了局部-全局流形结构保持的多视图关联学习方法，建立了亚像素-像素-超像素多视图协同学习分类框架，提升了标注样本稀缺情况下的高光谱图像整体分类精度。该项目研究成果为高光谱遥感图像高速高精度地物分类提供了新思路，为高光谱遥感在多个领域的广泛应用提供有力的技术支持。