NASA's satellites currently collect hundred of gigabytes of data and images per day as they orbit Earth. This volume of data is beyond the capacity of scientists to manually analyze the data stream, requiring automated methods of data processing, analysis, and discovery.

As part of our research into Intelligent Ecosystem Prediction with Identification and Analysis of Extreme Events (IEP), NASA Ames is collaborating with researchers at Carnegie Mellon University (CMU) and the University of West Florida Institute of Human and Machine Cognition (IHMC) to develop new techniques for processing these large volumes of data and apply machine learning techniques to discover new causal relationships in the data streams.

While there are many data sources we could use to develop and test the machine learning techniques, we have selected fire occurence as our initial demonstration problem. Not only is fire an important biospheric event which can have devastating impacts for local communities, but it is also a process which occurs at a scale that is detectable by satellites. While the occurrence of individual fires is in part a stochastic process, the development of conditions favorable for fire is a process which can be observed, modeled, tracked, and eventually predicted.

To date, we have retrieved, produced, and integrated a variety of heterogeneous data sources to produce a 20-year dataset for the continental U.S. of measurements related to fire risk. The dataset was generated by TOPS and the IMAGEbot Planner and includes TOPS 8km data products for the U.S. in addition to NASA's MODIS data products, and DAYMET and RUC meteorological data. Inputs which could not be not be obtained directly from satellite data or other sources (e.g., soil moisture, evapotranspiration, snow cover, and gross primary production) were produced by TOPS using satellite data as inputs.

Researchers at CMU and IHMC have been mining this data using machine learning techniques and have recently developed a number of promising causal models for fire risk for the data. We are currently in the process of integrating these models into the Ecocasting architecture for testing and validation using current and historic data sources. We plan to begin producing weekly and seasonal fire risk maps for the continental U.S. in Summer, 2004, so please check back soon for updates.