IEEE Geoscience and Remote Sensing Magazine - March 2013 - 49

NLCD Classification Scheme
Landsat 5 Bands

Classification Algorithm

Image Classification Seeks to Use Satellite
Imagery Data to Identity or Classify the Features
(Trees, Grass, Road, etc.) on the Ground.
NLCD (National Land Cover Data) Maps
Gathered in 1992 and 2001 by the MultiResolution Land Characteristics Consortium
(MRLC) Are Used As Truth Data for
Classification.
IMAGESEER Provides Not Only Selected
Satellite Data but Geo-Registered Ground Truths
or Benchmarks for Image Classification and
Cloud Cover Detection.

Image Classification

FIGURE 3. Example of validation data for the classification of Landsat data.

from the projects, e.g., the Automatic Cloud Cover Assessment (ACCA) program [14] for all Landsat-7 datasets, and
geo-registration is performed using methods developed
by our team [1], [15].
In terms of classification ground truth, NLCD classification maps are available for 1992, 2001, and 2006. Since
2006 were not available when ground truth maps were
being first collected for IMAGESEER, only 1992 and 2001
have been used. Image data were chosen around those
dates, within a 5-year time frame of these 2 dates, from
1990 to 1994 and from 1999 to 2003, respectively. The
1992 maps are used as classification truth data for scenes
from 1990 to 1994 and the 2001 maps are used as classification truth data for scenes from 1999 to 2003. This
assumes that changes in land cover were minimal within
a 2-year timeframe around the land cover acquisition
time. Figure 3 summarizes this concept of benchmark
data for image classification, showing the original Landsat 5 bands, the corresponding NLCD map and the classification scheme used by MRLC to develop these maps.
Additionally, Figure 3 shows a pie chart representing the
distribution of the various classes in the given dataset; this
feature is available on the website and will be described in
section IV.
march 2013

ieee Geoscience and remote sensing magazine

In terms of cloud cover, the ACCA algorithm was developed for Landsat-7 datasets, but we are currently investigating how to adapt this algorithm and others developed
for MODIS and EO-1 to complete the cloud cover maps for
all datasets.
Along with the
In terms of geo-registration,
all datasets are currently being
original data,
registered using an algorithm
associated validation
described in [15] and using
data are provided,
wavelet features and a Levene.g., cloud cover
berg-Marquardt optimization
masks and
to register all sensor data to the
assessments,
corresponding NLCD classificageo-registered
scenes,
tion maps. We are also currently
and classification
considering adding equivalent
datasets extracted from the
maps from the
Global Land Survey (GLS) that
National Land Cover
exist for 1990 and 2000 [16].
Data (NLCD).
In terms of gap filling, Landsat 7 datasets from the Enhanced
Thematic Mapper Plus (ETM+)
are currently being collected. After the failure of the Scan
Line Corrector (SLC) in 2003, ETM+ continued collected data
49

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