Semantic indexing and temporal rule discovery for time-series sattelite images

Feature extraction and knowledge discovery from a large amount of image data such as remote sensing images have become highly required recent years. In this study, we present a framework for data mining from a set of time-series images including moving objects using clustering by self-organizing mapping(SOM) and extraction of time-dependent association rules. We applied this method to weather satellite cloud images taken by GMS5 and evaluated its usefulness. The images are classied automatically by two-stage SOM. The results were examined and the cluster addresses were described in regard to season and prominent features such as typhoons or high-pressure masses. Sequential images are then transformed into a data series expressed by cluster addresses and time of occurrence, from which timedependent association rules (simple serial rules) are extracted using a method for nding frequently co-occurring term-pairs from text. Semantic indexed data and extracted rules are stored in the database, which allows high-level queries by entering SQL through user interface, and thus supports knowledge discovery for domainexperts. We believe that this approach can be widely useful and applicable to knowledge discovery from an enormous amount of multimedia data, which includes unknown sequential patterns.