Geoscience Reference
In-Depth Information
quite variable (i.e., metric dependent), which complicates assessment efforts (Hess, 1994; Shao et
al., 2001).
A major obstacle to assessing the accuracy of LC maps is the high cost of generating reference
data or multiple thematic maps for subsequent comparative analysis. Commonly employed solutions
include (1) selecting subsectional maps from a region (Riitters et al., 1995), (2) subdividing regional
maps into smaller maps (Cain et al., 1997), or (3) creating multiple maps using computer simulations
(Wickham et al., 1997; Yuan, 1997). Maps created using the first or second method are spatially
incompatible or incomparable, while maps created using the third method contain errors that do
not necessarily represent those found in actual LC maps. Therefore, it is necessary to create multiple
maps for a specific geographic area using different analysts or different classification methods
(Shao et al., 2001). The approach presented here represents an actual image data analysis and,
therefore, conclusions drawn from the analysis should be broadly applicable.
Past studies have focused on only a few indices. Hess and Bay (1997) made a breakthrough in
quantifying the uncertainties of adjusted diversity indices. Various statistical models have also been
developed to assess the accuracy of total area (%LAND) for individual cover types (Bauer et al.,
1978; Card, 1982; Hay, 1988; Czaplewski, 1992; Dymond, 1992; Woodcock, 1996). However, few
have used modeling to perform area calibrations (Congalton and Green, 1999). Shao et al. (2003)
derived the Relative Area Error (REA) index, which has causal relationships with area estimates
of LC categories. This study employed multiple classifications and reference maps to demonstrate
how classification accuracy affects landscape metrics. Here the overall accuracy and REA were
compared and a simple method was demonstrated to revise %LAND values using corresponding
REA index values.
15.2 METHODS
Multiple thematic maps were derived from subscenes of Landsat Thematic Mapper (TM) data
for two sites (A and B) located in central Indiana and the temperate forest zone on the eastern
Eurasian continent (at the border of China and North Korea). LC mapping was performed to
approximate a Level I classification product (Anderson et al., 1976). Site A thematic maps included
the following classes: (1) agriculture (including grassland), (2) forest (including shrubs), (3) urban,
and (4) water. The second site included only forest and nonforest (clear cuts and other open areas)
cover types. A total of 23 independent thematic maps were developed for site A. Analysts (
= 23)
were allowed to use any method to classify the TM imagery acquired on October 5, 1992. LC maps
were evaluated based on the overall accuracy. All the accuracies were comparable because all
assessments were performed using the same reference data set. Students performed the image
analysis, thus representing work performed by nonprofessionals (Shao et al., 2001).
Eighteen thematic maps were created for site B using a single TM data set acquired on
September 4, 1993, and a stack data set combining the 1993 data with other TM data acquired on
September 21, 1987. Training samples were acquired using three methods, including (1) computer
image interpretation, (2) field observations, and (3) and a combination of the two. Three classifi-
cation algorithms were used, including (1) the minimum distance (MD), (2) maximum likelihood
(ML), and (3) extraction and classification of homogeneous objects (ECHO). Our goal was to make
the classification process repeatable, and therefore to represent a professional work process (Wu
and Shao, 2002). Two additional maps with 94.0% and 94.5% overall accuracy that were created
with alternative approaches were also incorporated into this study.
The overall accuracy of these
maps ranged from 82.6% to 94.5% (Wu and Shao, 2002). More importantly, a reference map was
manually digitized for site B. The errors of landscape metrics of each map were computed as:
n
E
=
(
I
-
I
) /
I
¥
100
(15.1)
index
map
ref
ref
