Geoscience Reference
In-Depth Information
a time and cost savings as well as allowing the assessment of these landscape features over extensive
geographic areas such as the Chesapeake Bay. As part of the Multi-Resolution Landscape Charac-
terization 2000 program (MRLC 2000, 2002), the United States Geological Survey (USGS) has
embarked on an effort to map impervious surfaces across the conterminous U.S. utilizing subpixel
techniques. This study proposes to produce a spatial and statistical framework from within which
we can investigate subpixel-derived estimates of a material of interest (MOI) utilizing multiple
accuracy assessment strategies.
Traditional map accuracy assessment has utilized a contingency table approach for assessing
the per-pixel accuracy of classified maps. The contingency table is referred to as a confusion matrix
or error matrix (Story and Congalton, 1986). This type of assessment is a “hit or miss” technique
and produces a binary output in that a pixel is either “correct” or “not correct.” The generally
accepted overall accuracy level for land-use (LU) maps has been 85% with approximately equal
accuracy for most categories (Jensen, 1986). While alternative techniques to assess the accuracy
of land-cover (LC) maps using measurement statistics such as the Kappa coefficient of agreement
have been proposed, most methods still rely on the contingency table and use per-pixel assessments
of the thematic map class compared to “truth” sample points (Congalton and Green, 1999). However,
as noted by Ji and Jensen (1999), this classic “hit or miss” approach is problematic with respect
to assessing the accuracy of a subpixel-derived classification. A subpixel algorithm allows the pixel
to be classified based on the percentage of a given MOI such that for any given pixel the “fit” to
truth can be assessed. A level of accuracy can be still be obtained from a pixel that “misses” the
truth. The derivation of a percentage of a MOI per-pixel allows for alternative accuracy assessment
approaches such as aggregate whole-area assessments (i.e., watershed) and correlations (Ji and
Jensen, 1999). These alternative approaches may produce adequate accuracies despite the fact that
a lower per-pixel accuracy is derived from the standard error matrix.
An accuracy assessment of subpixel data is largely dependent upon high-resolution planimetric
maps or images to provide reference data. Concurrent with the emergence of subpixel techniques
has been a trend in the production of high-resolution data sets, including high-resolution multispectral
satellite imagery, GIS planimetric data, and USGS Digital Ortho Quarter Quads (DOQQs). All these
data sources can be readily processed within standard GIS software packages and used to assess the
accuracy of subpixel estimates, as derived from Landsat data, over large geographic regions.
In this study we compared classified subpixel impervious surface data derived from Landsat
TM imagery and planimetric impervious surface maps produced from photogrammetric mapping
processes. Comparisons were performed on the classified subpixel (30 m) using planimetric refer-
ence data in a raster GIS overlay environment. Our goal was to produce a spatial framework in
which to test the accuracy of subpixel-derived estimates of impervious surface coverage. In addition
to a traditional per-pixel assessment of accuracy, our technique allowed for a correlation assessment
and an assessment of the whole-area accuracy of the impervious surface estimate per unit area (i.e.,
watershed). The latter is important for ecological and water quality models that have percentage
of impervious surface as a variable input.
19.2 METHODS
19.2.1
Study Area
subwatershed located 9 km west
of Baltimore, Maryland (Figure 19.1). The Dead Run subwatershed is a portion of the greater
Baltimore Long Term Ecological Research (LTER) area located in Baltimore County, Maryland,
and resides within the coastal plain and piedmont geologic areas of the Mid-Atlantic physiographic
region. Previously produced planimetric and subpixel data sets were available for the area.
Our study area was the Dead Run watershed, a small, 14-km
2
