多时相遥感变化检测的动态贝叶斯网络研究

Study on Dynamic Bayesian Networks for Multi-temporal Remote Sensing Change Detection

动态贝叶斯网络是20世纪90年代在贝叶斯网络基础上发展起来的、利用时序动态数据产生可靠概率推理的新方法,动态贝叶斯网络为实现遥感变化检测从静态到动态分析提供了一种新的途径。在实现了贝叶斯网络遥感数据分类的基础上,把握发展动态,探索了利用动态贝叶斯网络对多时相多特征遥感数据进行变化检测的问题。以北京东部地区1994年、2001年和2003年5月Landsat TM遥感数据为例,介绍了利用动态贝叶斯网络进行多时相遥感变化检测的基本过程。实验结果表明:动态贝叶斯网络算法可以一次性输入和处理多个时相的遥感数据,并通过概率和有向无环图表达了不同时间片段之间特征和状态变化的关系。

The Dynamic Bayesian Network （DBN）, which uses the time-series dynamic data to produce credible probabilistic reasoning, is a method developed in 1990s based on the Bayesian network, and offers a way to change analysis from the static viewpoint to the dynamic viewpoint when we carry out remote sensing change detection. Grasping the development tendency, we explore how to use Dynamic Bayesian Networks for direct change detection of remote sensing data with multi-temporal features. Taking the Landsat TM remote sensing data of eastern Beijing area acquired in May of 1994, 2001 and 2003 as an example, we introduce in detail the method to do multi-temporal remote sensing direct change detection using Dynamic Bayesian Networks. The good result indicates that： the DBN-based direct change detection algorithm can input and handle remote sensing data of more than two time phases simultaneously, and it describes the relationship among the features and states of different time phases by means of probability and directed acyclic graphs.