Temporal stability of an ERS-1/JERS-1 SAR classifier

With the launches of the European ERS-1 and Japanese JERS-1 satellites worldwide SAR data is now routinely available from a spaceborne sensor. The combination of these two sensors promises even more discrimination ability than either alone. Consequently, it is important to devise a classification algorithm that is robust to variability in seasonal variations, weather, and local vegetation species. This paper presents a first step toward that goal. Several scenes acquired during August, and one during December using ERS-1 were combined with two scenes from August and October from JERS-1, but from different years. Each same-season pair is then combined and classified using a hierarchical Bayesian approach. The August ERS-1 scenes were chosen based on local moisture conditions: some having relatively dry soil and another scene was acquired just after a significant rain storm. The October/December pair is used to assess the ability to classify in the fall/winter. Despite the variability in conditions, a single classifier is to be developed that can classify five or six structural classes (bare surfaces, short vegetation, and a few different kinds of trees) with high accuracy. This has been achieved with a single ERS/JERS image with accuracies higher than 90%. Once all the different scenes have been classified the differences between the rules in each of the classifiers is to be related to changes in physical parameters due to rain and the change of season. For example, since the deciduous trees do not have leaves during the winter, the C-band (ERS-1) radar response is quite different in the summer than in the winter. Basic knowledge such as this can be used to adapt the classifier to seasonal changes