空间逐步寻优的数据挖掘法的多波段影像分类研究

对于在特征空间中寻找特征模式,一般是通过假设分布函数一次性对样本空间进行分离的方法去试图获得特征空间的样本总体分布规律。但是由于样本集之间互相重叠或者由于离散样本相互干扰的原因,往往很难获取细节性和过程性的分布结构,而直接影响结果的精度和解释力。本文提出了空间逐步寻优的数据挖掘方法(SOMM),是在遗传算法寻优理论基础上,根据知识参数化样本分布函数,来逐步分离样本空间,获得样本空间的树状的层状分布结构;同时提出了基于SOMM的多波段遥感影像聚类模型和监督分类模型;最后对分类过程和结果进行了综合分析,通过与最大似然和BP神经网络方法相比较,认为SOMM方法在过程化、细节化、分类精度、融合领域知识等方面具有一定的优势。

Stepwise Optimization Making Modal for Spatia Data Mining and Its Application in Multi Spectral Remote Sensing Image Classification

In order to find out the feature patterns from multi-dimension space, the conventional approach is to separate feature space by assuming the distributed functions of all features in one time. But however, because of inter-overlapping among the sample sets and the confusion from the discrete points, it is often very difficult to acquire the subtle and procedural distribution structure, which can influence the precision and interpretability of the mining outcome. This article presents a new modal named Stepwise Optimization Making Modal (SOMM) which is used to dig out the tree-like hierarchical distribution structure from feature space based on the Genetic Algorithms Optimization Theory and the knowledge fused distribution functions. The SOMM based multi-spectral remote sensing classification approaches, including the hierarchical clustering modal and supervised hierarchical classification, are also presented. Finally, the case of practical application of remote sensing land covering classification in Hong Kong region is presented, then the procedure of SOMM and conventional MLC approaches are synthetically analyzed. Experimental results show that SOMM approach has more advantages in procedural prediction, subtle structure, classification precision, and area knowledge fusion, etc.