Automated event classification for multi-gigabyte per day data streams

Multiband and hyperspectral scanning and staring remote sensors can generate dozens of gigabytes of data each week. Manually searching through so much data for short-lived, transient events is prohibitively expensive. The goal of our data exploitation program is to develop real-time processing algorithms to identify and classify such events for data streams of this magnitude. In this paper we describe a Bayes classifier that can assign identified events to known broad categories and can designate a confidence for correct classification based on the measurement conditions. Events that do not fit known categories are marked for the analyst´s attention. We have developed a graphical user interface to allow an analyst to process large numbers of events at a time. Our data exploitation program has developed classifiers in addition to those discussed here. When classifiers working in parallel disagree on the assignment of an event to a category, some means is needed to decide which to choose. To make the decision, we have devised a “Committee of Experts” approach that arrives at a final classification by taking into account the confidence of the individual classifiers under the given measurement conditions