IEEE Geoscience and Remote Sensing Magazine - March 2013 - 46

ing 7 bands whose wavelength varies from the visible
to the thermal infrared range (in comparison, a typical
high-definition computer image is 1920x1080 pixels
with 3 "bands": red, green, and blue). Hyperspectral
data can be even worse with hundreds of bands per
pixel. Handling such amounts of data in real-time must
take into account computational requirements such as
speed and memory. As a consequence, the implementation of such methods on parallel, distributed or even
on-board computers must be considered.
◗◗ The lack of a known image model. Similarly to fiducial
points, a very rough sketch of a city, containing a river or
a network of roads, can be utilized as a global model for
Earth Science scenes; or a geological map can be utilized
for planetary data. However, this usually lacks either
the amount of detail and/or the degree of invariance to
acquisition conditions (e.g., atmospheric and seasonal
variations) that would be needed for optimal and reliable image processing.
◗◗ The lack of well-distributed "fiducial points" or natural "truth
data" resulting in the difficulty to validate image processing methods in the NASA domain. For example, for georegistration of Earth Science data, it is possible to use
well-known man-made landmarks as fiducial points,
but such landmarks are very rare, and are not evenly
distributed around the globe. The key factor in any accuracy assessment of satellite data is linked to the ability to gather ground reference data independently of
the satellite data themselves. Additionally, depending
on the time between the on-site
ground reference gathering and
the imaging of the area, the vaIMAGESEER provides
lidity of the ground reference
a widely accessible
data may be lessened due to
database of
anthropomorphic or natural influences. Another approach is to
NASA-centric,
compare the digital image with
easy to read image
other sources of ground referbenchmark data.
ence data, such as other instruments data or appropriate reference maps, provided that the
features of interest are detectable using these sources.
The degree of correspondence between the ground reference data and the measurement derived from the sensor
data can then be compared for accuracy assessment.
C. Existing Databases Containing NASA Data
Various databases already exist that provide benchmark
image datasets for teaching or validating image processing algorithms [2]-[5]. Including different types of test
images (e.g., medical, military, fingerprint, pedestrian,
leaf, etc.) with very few NASA images, they represent various image processing challenges, such as motion detection,
image segmentation, object recognition, etc. Despite these
various resources, very few of these test datasets represent
NASA-centric applications and none of those databases
46

contain the knowledge or validation information necessary
to assess the value of an algorithm for the specific NASA
application. A few other databases [6]-[11] focus on Earth
remote sensing data and space data and can be classified
into two types:
1) Large repositories such as the NASA Distributed Active
Archive Centers (the DAAC's), the University of Maryland Global Land Cover Facility (GLCF), the Planetary
Data System (PDS) and the USGS Earth Resources Observation and Science (EROS) Center [6]-[9]. Those
data collections, aimed at Scientists working in Earth
or Space Science, usually require a great deal of background knowledge in order to search, read, and interpret
the data.
2) Smaller image data collections such as the NASA Image
Exchange (NIX, [10]) or the Jet Propulsion Laboratory (JPL) Photojournal [11] focusing on some wellselected and illustrative images from various NASA domains. Although closer in spirit to the proposed focus
of IMAGESEER (smaller amount of data, easy formats
and educational material), these data collections represent "image galleries" used for public outreach, or
for illustrative purposes, and they cannot be used as
benchmark datasets.
For more details on those various databases, see [12].
III. IMAGESEER Technical Description
A. Overall Description
The IMAGESEER concept was developed under the
NASA Goddard Space Flight Center's Internal Research
and Development (IRAD) program and was designed
in collaboration with both the NASA Science and the
University communities. A first prototype is available
at http://imageseer.nasa.gov. As shown in our first prototype, IMAGESEER uses a graphically-rich web site for
easily browsing and downloading all of the selected datasets, benchmarks, and tutorials. Using a paradigm common
on commercial websites, users can restrict their searches by
selectively filtering by data source (project, mission, and
instrument), region of interest, desired image processing
technique, and time period. Figure 1 shows IMAGESEER
web user interface front page.
Compared to other larger data repositories such as the
DAAC, the GLCF or the PDS, the IMAGESEER database is
deliberately much smaller and focused in scope, the data
is already pre-processed for ease of use (e.g., is the data correctly geo-registered to the correct latitude/longitude), the
size and the format of the datasets are user-friendly and
most importantly, all datasets have corresponding "truth
data" (e.g., the "answers" to applications of interest such
as a validated classification image identifying every pixel
(forest, farmland, etc) in a scene) and application-related
domain information. The website has also small educative introductory material (called Imagepedia) for subjects
such as remote sensing, image processing, Land Cover/
ieee Geoscience and remote sensing magazine

march 2013

http://imageseer.nasa.gov

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2013