Knowledge Discovery from Database and Its Application in Remote Sensing Inversion

摘要

With the advent of Earth Observation Program, the speed of data accumulation and update is getting faster than ever. However, the gain of the knowledge from the data has not been progressed as what is expected while the data volume is getting immensely huge. Knowledge discovery from database (KDD) is a feasible method to solve this contradictive problem of "the lack of knowledge while data getting explosive". In addition, the uncertainty in the Remote Sensing Science is worth of being investigated as well due to the fact that the remotely sensed data may be collected under various circumstances, which resulted in multi-sourced information. The process of acquiring the data is another factor impacting the data quality, because the transmission of the signal is affected by all kinds of uncertain factors. Thus, uncertainty is an inherent property of the Remotely Sensed data(Michel Crosetto, 2001).rnAs a graphical model for the probabilistic relationships among a set of variables, the Bayesian network has become a popular representation for encoding uncertain expert knowledge in KDD domains over the last decade(Datcu and Seidel, 1999; Marcot et al., 2001; Murphy, 1998). The Bayesian method integrates a prior knowledge about the objects under study and the information provided by new data set, followed by encoding the multi-knowledge into conditional probability network model. Thus, Bayesian network in conjunction with Bayesian statistical techniques facilitates the combination of domain knowledge with the relevant data. Theoretically, the updated information is weighted by the prior knowledge together with the information induced by the new data, and the uncertain knowledge has been considered as the probabilistic causal relationship among parameters(Chan and Darwiche, 2005).rnThis paper investigates the causal relationship between Bidirectional Reflectance Distribution Function(BRDF) and some other earth surface status parameters, such as the type of land objects, the temporal factors of vegetation growing(e.g. phenology period) and the planting structural parameters(e.g. Leaf Area Index(LAI), Leaf Angel Distribution(LAD))(Li et al., 2001; Marie Weiss, 2000). By learning the Bayesian network parameter from the database, the associated knowledge on the reflectance on characteristic band of BRDF and the earth surface parameter are established. This type of knowledge can be used as the constrained factor named soft-bound in land surface parameters inversion algorithm. As illustrated, using KDD technique under Bayesian network to discover the uncertain knowledge from Remotely Sensed database can help accumulate the prior knowledge and support the application of vegetation parameters inversion.