Bayesian feature selection for classifying multi-temporal SAR and TM data

Remotely sensed imagery data from various satellite sensors are now available for environmental monitoring. However, due to the difficulty in surveying, it is not easy to obtain a sufficient number of training data for classifying these high dimensional imagery data. In order to make use of these imagery data, it is necessary to develop a classification method which can attain a high classification accuracy only using a limited number of training data. In this study, the authors have tested the Bayesian approaches which integrate feature selection and model averaging in the classification process. The experiments are conducted using bayesian neural networks, Gaussian process, and maximum likelihood for classifying wetland vegetation types using multi-temporal LANDAT/TM, JERS1/SAR, and ERS/SAR data. The results shows that the Bayesian approaches work well for classifying these imagery data, and especially the Gaussian process has a very high accuracy which outperforms other methods for classifying the sensor fusion data using JERS1/SAR and LANDSAT/TM