面向土地利用分类的多源遥感数据混合贝叶斯网络分类器

A Hybrid Bayesian Network Classifier for Multi-source Remote Sensing Data in Land Use Classification

传统的离散型贝叶斯网络分类器是将所有变量视为离散变量,或对连续变量做离散化处理。可是离散化不可避免地存在信息损失,且在多源遥感数据的处理和分析中,连续变量的离散化会导致搜索空间的急剧增加和计算及存储量的极大开销。针对这些问题,开发了一种面向土地利用分类的多源遥感数据混合贝叶斯网络分类器,该分类器首先对问题领域的所有变量做正态分布检验,同时将满足高斯分布假设的变量不做离散化而视为连续变量;然后对离散变量和连续变量分别进行参数学习,最后进行参数合并,再用于贝叶斯网络的推理和分类中。通过福州市区土地利用分类的实验表明,该模型优于传统的离散型贝叶斯网络分类器,具有一定的研究和应用价值。

It is necessary that all variables be considered as discrete variables, or discretization be conducted in a traditional discrete Bayesian network classifier. The information loss in discretization is inevitable, and the discretization of continuous variables will lead to dramatic expansion of search space and great expenses in computation and storage in multi -source data processing and analysis. To solve these problems, the authors have developed the Hybrid Bayesian network classifier for land use classification, which first conducts normal distribution test for all variables in the study area. For the variables that meet Gaussian distribution assumptions, the authors do not discrete them and regard them as continuous variables. Parameter learning of discrete variables and that of continuous variables are carried out respectively, and then the parameters are merged. These parameters are used for reasoning and classification of Bayesian network at last. Experiments of land use classification in Fujian show that the model is superior to the traditional discrete Bayesian network classifier, and hence has great research and application value.