Estimating and Mapping the Thematic Accuracy
of GAP Land Cover Maps
ROLAND L. REDMOND
Montana Cooperative Wildlife Research Unit, University of Montana,
In keeping with the research and development mission of the Gap
Analysis Program, we developed an approach to accuracy assess-
ment for land cover in Montana that did not require the collection
of an independent set of reference data. Instead, all available refer-
ence data were used not only to train supervised image classifica-
tions, but also to assess the resulting classification accuracies via a
bootstrap procedure. Moreover, because standard classification error
matrices provide little information about the spatial variation in the-
matic accuracy, we used Kriging to interpolate probability estimates
from each reference point to a statewide lattice, from which a con-
tour map of thematic accuracy was produced. Although the method
has withstood peer review (Steele et al. 1998), no results have been
field-tested, and readers should be cautious in applying it to their
areas or states. For more details about the procedure and how to
construct bootstrap classification error matrices, see Steele et al.
(1998); additional details about the statewide application can be
found in Redmond et al. (1998).
The Montana Gap Analysis Project (MT-GAP) began before the
widespread use of airborne videography. Because of the large size
of the state and a lack of funds to collect new ground-reference
data, we were limited to the use of whatever existing data were
available. Although this amounted to 21,348 plots representing 45
cover types, this was not a sufficiently large sample for us to hold
back a certain proportion (e.g., 20%) to validate the supervised clas-
sifications of 33 TM images covering the state. Thus, we devised a
method of estimating the probability of misclassification at each of
these reference points using a bootstrap procedure (Efron and
Tibshirani 1993). This method simulated the process of sampling
and classification many times (with replacement), and thereby al-
lowed us to estimate the probability that the true cover type was
correctly classified at each reference point from the number of times
that the reference observation was correctly classified in all the simu-
lations. For MT-GAP, we ran the bootstrap 100 times with replace-
ment for each TM image classification. The resulting probability
estimates then were entered into ARC/INFO (GRID module), and
mean thematic accuracy was interpolated to a 1 km statewide lat-
tice using the routine POINTINTERP (exponential option with
neighborhood = 75 km, and decay = 15 km); in other words, mean
accuracy was calculated for each lattice point across the state using
only the reference data (cover types and accuracy estimates) that
fell within a 75 km search radius. Finally, contour lines connecting
lattice points of equal mean thematic accuracy were drawn at 5%
intervals. Not surprisingly, the resulting map showed considerable
spatial variation in mean thematic accuracy for the 45 cover types.
Yet despite the utility of these results, we acknowledge that an in-
dependent validation (as recommended by Stoms et al. 1994 and
Crist and Deitner 1998) would still be worthwhile, especially to
detect and measure errors resulting from the possible omission or
inadequate representation of cover types in the sample of existing
Crist, P., and R. Deitner. 1998. Assessing land cover map accu-
racy. Version 2. Gap Analysis Handbook.
Efron, B., and R.J. Tibshirani. 1993. An introduction to the boot-
strap. Chapman and Hall, New York.
Redmond, R.L., M.M. Hart, J.C. Winne, W.A. Williams, P.C.
Thornton, Z. Ma, C.M. Tobalske, M.M. Thornton, K.P.
McLaughlin, T.P. Tady, F.B. Fisher, and S.W. Running. 1998.
The Montana Gap Analysis Project Final Report. Unpublished
Steele, B.M., J.C. Winne, and R.L. Redmond. 1998. Estimation
and mapping of misclassification probabilities for thematic land
cover maps. Remote Sensing of the Environment 66: 192-202.
Stoms, D., F. Davis, C. Cogan, and K. Cassidy. 1994. Assessing
land cover map accuracy for Gap Analysis. Version 1. Gap