贝叶斯网络支持的地表参数混合反演模式研究  

Study on Hybrid Inversion Scheme under Bayesian Network

作　　者：屈永华[1] 王锦地[1] 刘素红[1] 万华伟[1] 周红敏[1] 林皓波[1] 

机构地区：[1]北京师范大学地理学与遥感科学学院环境遥感与数字城市北京市重点实验室遥感科学国家重点实验室,北京100875

出　　处：《遥感学报》2006年第1期6-14,共9页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家高技术研究发展计划(863计划)(2002AA130010);国家重点基础研究发展规划项目(G20000779);国家自然科学基金项目(40571107);高等学校博士学科点专项科研基金(20040027019)资助

摘　　要：基于贝叶斯网络理论,建立用于植被地表参数估计的混合反演模式,结合遥感物理模型实现了冬小麦叶片叶绿素含量(Cab)和冠层叶面积指数(LAI)的反演。用模型模拟数据以及2001年顺义遥感实验数据验证结果表明,LAI和Cab均有较好的反演精度。针对含噪声模拟数据反演结果中约有10%的噪声数据反演失败的情况,用不确定知识的处理方法有效地降低了失败点的比例。混合反演模式本质上是一个融合先验知识与观测数据的知识推理方案,本文实现了对反演过程中参数后验概率更新算法并引入热力学中的信息熵概念实现了参数后验信息动态定量计算,同时简单探讨了现阶段定量评价遥感反演过程中信息流控制存在的难点问题。A hybrid inversion scheme for estimating surface variables of vegetation is proposed under Bayesian Network（BNet） theory, and then is used to estimate chlorophyll content of winter wheat leaves（Cab） and Leaf Area Index（LAI） of canopy. A coupled physical model named PROSPECT ＋ SAIL was chosen to generate simulation data set, which means that the SAIL model uses the leaf reflectance and transmittance derived from PROSPECT model to simulate canopy directional reflectance. Results derived from simulation data and SHUNYI Experiment in 2001 data show that both LAI and Cab can be estimated with an appreciated accuracy under the proposed scheme, except that there are about 10% of total points falling into failure inversion. Then an uncertain data handling method, which considers the measured data as the random variables obeying Gaussian distribution, is employed to solve the failure problem. As a resuh the failure points are removed successfully though the RMSE of estimated the two variables is larger slightly. The presented hybrid inversion scheme is a knowledge-based inferring mechanism in principle, so the updated information content in the inversion process is quantitatively calculated thanks to the concept of entropy introduced from thermodynamics. Contrasting to the conditional entropy, the posteriori entropy calculated according to our proposed probability revision algorithm is not a descending parameter. This property can give some indications in estimating the information content parameters and the currently used data, that is to say, if the data are consistent with the previously derived information of estimated parameters, then there is descending entropy, otherwise, it is ascending. In the last section of this paper, some discussions are presented about the problem on how to estimate and control the information stream, especially when the inversed physical model is nonlinear.

关 键 词：贝叶斯网络 混合反演 波谱库 信息熵 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]