Extending TOPS: A Prototype MODIS Anomaly Detection Architecture

The management and processing of Earth science data has been gaining importance over the last decade due to higher data volumes generated by a larger number of instruments, and due to the increase in complexity of Earth science models that use this data. The volume of data itself is often a limiting factor in obtaining the information needed by the scientists; without more sophisticated data volume reduction technologies, possible key information may not be discovered. We are especially interested in automatic identification of disturbances within the ecosystems (e,g, wildfires, droughts, floods, insect/pest damage, wind damage, logging), and focusing our analysis efforts on the identified areas. There are dozens of variables that define the health of our ecosystem and both long-term and short-term changes in these variables can serve as early indicators of natural disasters and shifts in climate and ecosystem health. These changes can have profound socio-economic impacts and we need to develop capabilities for identification, analysis and response to these changes in a timely manner. Because the ecosystem consists of a large number of variables, there can be a disturbance that is only apparent when we examine relationships among multiple variables despite the fact that none of them is by itself alarming. We have to be able to extract information from multiple sensors and observations and discover these underlying relationships. As the data volumes increase, there is also potential for large number of anomalies to "flood" the system, so we need to provide ability to automatically select the most likely ones and the most important ones and the ability to analyze the anomaly with minimal involvement of scientists. We describe a prototype architecture for anomaly driven data reduction for both near-real-time and archived surface reflectance data from the MODIS instrument collected over Central California and test it using Orca and One-Class Support Vector Machines algorithms. We demonstrate our efforts in the context of the Terrestrial Observation and Prediction System (TOPS), a flexible modeling software system that integrates ecosystem models with frequent satellite and surface weather observations to produce ecosystem nowcasts (assessments of current conditions) and forecasts useful in a range of applications including natural resources management, public health and disaster management.