基于多层次分割分类模型及其特征空间优化的建筑物提取方法

Extraction of buildings in remote sensing imagery based on multi-level segmentation and classification hierarchical model and feature space optimization

针对高空间分辨率遥感影像城市地物信息提取中的尺度效应、光谱多样性及分类特征优化等问题,基于面向对象影像分析方法,结合数据挖掘与机器学习技术,提出了一种多层次分割分类模型及其特征空间优化的建筑物提取方法。首先,根据遥感信息多尺度特性,针对地物特征差异设立层级关系,再结合光谱多样性特征定义地物包含的子类型,建立基于不透水面分割分类提取建筑物的层次化结构;然后,利用提出的ReliefF-PSO组合特征选择方法,优化构建相应层次的影像特征空间;最后,在多尺度分割和特征优化的基础上,基于随机森林模型获取不透水面分布,进而采用J48决策树算法分类提取建筑物。实验结果表明,该方法能够利用较少数量的影像特征,获得高精度的建筑物提取成果。

In view of the problems of scale effect, spectral diversity and classification feature optimization in the extraction of urban objects information from high spatial resolution remote sensing images,the authors, based on the object-based image analysis method and combined with data mining and machine learning,propose a multi-level segmentation and classification hierarchical model and its feature space optimization method for building extraction. First, according to the multi-scale characteristics of remote sensing information, a hierarchical relationship is set up for the difference of features of ground objects, and then a hierarchical structure based on information segmentation and classification is established based on the characteristics of spectral diversity to define the subtypes of ground objects. After that, the proposed Relief F-PSO combination feature selection method is used. Finally,on the basis of multiscale segmentation and feature optimization, the water surface distribution is obtained based on the random forest model, and finally the building information is extracted by the J48 decision tree algorithm. Experimental results show that the method can utilize a small number of image feature attributes to get high-precision building extraction results.