窄幅高光谱图像辅助宽幅多光谱图像地表精细分类研究

An important work of land resource survey is to acquire the land cover and land use change information. Remote sensing technique has macro, comprehensive, and dynamic characteristics, it has played an important role in the land resource survey. Although the multispectral remotely sensed data are widely used and has become one of the main data sources, the existence of mixed pixels and the restricted features limit the effective use of multispectral data. However, the hyperspectral image applied in the resource investigation over a large area usually suffers from a problem of narrow swath. This study tries to combine the hyperspectral image (narrow swath) and multispectral image (broad swath) together and improve the multispectral image classification accuracy. The key points are how to reconstruct hyperspectral image from multispectral image, and how to extract training samples and effective classification features by the help of hyperspectral information. A practical classification method will be finally proposed to improve the classification accuracy based on the combination of multispectral and hyperspectral data. Obviously, this method will provide scientific foundation to support the investigation and monitoring of resources and environment over a large area, and will provide accuracy mapping result to help decision making about sustainable economic development.

土地资源调查中一项重要工作就是获取地表覆盖和土地利用变化的信息，遥感具有宏观、综合、动态、快速等优点在土地资源调查中发挥了重要的作用。虽然多光谱数据目前是土地资源调查的主要数据源之一，但是其存在的同物异谱、异物同谱、分类特征有限等问题却在一定程度上限制了有效应用。而高光谱数据因为幅宽通常较窄，在大范围资源环境调查和监测中也大大受限。本项目尝试以窄幅高光谱图像和宽幅多光谱图像的复合利用为研究点，探讨两种数据协同分类的途径和方法。希望借助条带高光谱图像的约束将全区多光谱图像重构为高光谱图像，鉴于高光谱数据的分析实现多光谱图像样本优选和特征优化这两种方式，实现窄幅高光谱图像和宽幅多光谱图像的复合利用。发展出一种少量高光谱数据辅助下可操作又能改善多光谱图像精细分类精度的分类方法，满足大区域高精度地表覆盖分类的实际应用需求，为大区域资源环境的调查和监测提供理论基础和依据。

针对高光谱数据光谱丰富但幅宽较窄的特点，项目提出了利用窄幅高光谱图像辅助宽幅多光谱图像进行精细分类的思路，并从样本优选和光谱重构两个角度开展了两者的复合分类研究。在样本优选方面，借助窄幅高光谱图像提取更为精细和纯净的训练样本，通过改善训练样本的代表性和精细化程度，提高了宽幅多光谱图像的精细分类水平。在光谱重构方面，提出了一种局部数据约束下，将区域多光谱图像重构为高光谱图像的方法，利用重构的高光谱图像挖掘和构建了新的分类特征，从而实现了大范围地表的精细分类。实验结果表明，少量高光谱数据的辅助可以有效改善大范围多光谱图像的分类精细化程度和水平，解决了大区域资源环境调查中高效与精细化分析之间无法平衡的问题，能够满足大区域高精度地表覆盖分类的实际应用需求。