|
|
|
PECASE: Applied Pattern Recognition Research
|
|
|
| NASA scientists are overwhwlmed by the vast amounts of data available
from instruments, experiments, and simulations, and all trends point to
continuing increases in the size of these datasets. Our ability to collect,
store, and process ever-greater volumes of data is guaranteed - but to
condense these data into scientific understanding, we will need new
analysis tools. We propose to enable pattern discovery in large
spatiotemporal datasets by developing new algorithms that find and analyze
objects, which we take to be spatially-coherent regions of interest that
evolve
through time. The notion of objects allows investigators to climb above a
pixel-level understanding of their data. These objects must therefore be found
in single images, tracked through image sequences, and their temporal
behaviours modeled quantitatively. In prior work, we have solved a large
class of object-identification problems with general statistical methods
that employ unsupervised learning or scientist-provided labelings to define
local assessments of activity. These pixel-level cues are linked into an
object-level scene description. Related time series models are used for
object tracking and analysis. Here, we propose to extend the underlying
methodology
to cope with much larger object sets, and to allow for non-Gaussian state
representations. This combination of capabilities will enable complex
scenes to be broken down in terms of interacting objects. The system is
field-tested on a vast multi-wavelength series of solar images from an
imager aboard SoHO. Strong collaborations with domain scientists ensure
successful deployment into other Sun-Earth connection missions, and NASA
application areas like planetarty science and geophysics. Expressing the
models in neutral, mathematical/statistical lanaguage allows
repeatability, model refinement over time, and model exchange among
investigators. Our use of open standards for structured data (e.g. object
models and datasets in XML with suitable schemata) ensures portability
across applications and extensibility. |
|
|
