基于条件随机场模型的多分辨率遥感图像复合分类算法

摘要

A novel compound classification algorithm for multi-resolution satellite data based conditional random field models is presented to improve the performance of land covering classification effectively by making use of the multiple spatial resolution satellite data at the arithmetic level. The approach is based on multiple data sources but not limited to full-scale high resolution data. The multi-to-single spatial correspondence is learnt from the sample area where the high resolution data is available. The nonparametric "real" likelihood distribution estimation is adopted and "real" likelihood features for low resolution pixels are extracted based on the cloud theory. The sequences of "real" likelihood features, which represent the relations between spectrum and land covering types, is integrated into the classifier with the spatial contextual information between pixels by defining two types of potential functions. The classifier based on conditional random fields offers a robust and accurate framework which can support multiple features and represents the special continuity of land covering. The experiments on the MODIS and TM satellite data show that the proposed method can greatly improve the accuracy for large scale land covering classification applications.%针对如何在算法层次上利用不同空间分辨率遥感数据提高地表分类精度的问题,提出了一种基于条件随机场模型的全新的多分辨率复合分类算法.该算法针对同一地区、不同覆盖范围的两种高低分辨率遥感图像,以广域低分辨率图像的高精度地表分类为目的,利用高低分辨率图像间的空间分辨率多对一关系,基于云理论构建“真实”似然特征映射,由用来描述光谱特征与类别关系的“真实”似然特征序列以及像元间上下文关系构建条件随机场模型的两类势函数,并在此基础上对广域低分辨率图像进行全局地表分类.该算法不仅提供了对多分类特征的支持,而且考虑了地物分布的空间连续性.多组高低分辨率图像组合下的复合分类及不同算法间的分类精度对比分析结果表明,该算法可有效提高广域低分辨率图像的分类精度,并具有良好的鲁棒性.