一种多源遥感影像非监督分类方法中的欠分割对象检测与拆分算法

Recognizing and Splitting Under-segmented Objects in An Unsupervised Classification Method Using Multi-source Remote Sensing Images

广义中餐馆连锁模型是一种基于全色和多光谱影像的非监督分类方法,它在一个非参数贝叶斯框架下同时实现基于全色影像的分割及基于多光谱影像的分类。由于全色影像光谱分辨率的限制,导致其所获取的部分分割体存在"欠分割"现象,影响模型最终分类精度。针对广义中餐馆连锁模型中的欠分割问题,提出基于广义中餐馆连锁模型的欠分割对象检测及拆分方法。首先,提出分割体的异质性指标以对可能包含多种地物的分割体进行检测;其次,基于多光谱影像得到的语义分割体提供的边缘信息对检测出的欠分割对象进行拆分;最后,基于多光谱影响完成分类。实验结果表明,改进后的模型能够有效减少广义中餐馆连锁模型基于全色影像获取的语义分割体的欠分割现象并提高模型分类精度。

Generalized Chinese restaurant franchise(gCRF)can be used for unsupervised classification using both panchromatic(PAN)and multi-spectral(MS)images.It integrates segmentation based on PAN images and classification based on MS images in a unified non-parametric Bayesian framework.However,the segmentation results obtained using PAN images have the problem of under-segmentation due to the low spatial resolution of PAN images,which reduce the accuracy of classification results.In this paper,we improve the gCRF by recognizing and splitting under-segmented objects to obtain better classification results.First,we propose a measure named heterogeneity index(HI)which is constructed using spectral values from MS images to screen the under-segmented objects derived from PAN images.After the under-segmented objects are selected,they are split to smaller segments each of which contains only one class according to the edge information provided by segments derived from MS images.The experiments show that the segmentation and classification results are both improved by introducing the process of recognizing and splitting the under-segmentation objects.