lated confidence for those pixels whose classification was determined based on two scenes. The
monotonic relationship between these variables confirms the earlier statistical arguments that
consistency is a legitimate “surrogate” of user's accuracy.
Multiple-scene LC products can be expected to exhibit significant internal variations in user
accuracy. Detailed characterization of this variability was not feasible using conventional ground
reference sampling because of cost and logistics. However, the level of interscene classification
consistency provided an indirect “surrogate” measure and was used to gauge local accuracy. This
alternative approach was especially attractive for application with Landsat-based maps since exten-
sive overlap areas exist for adjacent orbital paths located in nonequatorial latitudes.
Consistency measures were effectively employed using a number of processing steps. First,
assessments were evaluated at the cluster level, thereby providing an estimation of performance at
the level of the labeling unit rather than only at the class level. Then, by analyzing the consistency
during the product generation phase, detection and correction of incorrectly labeled clusters was
accomplished prior to the creation of the final product, thereby improving its quality. Finally, within
the interscene overlap regions, consistency served as a “compositing” criterion to select an optimum
label and could be accumulated to encapsulate the added confidence associated with multiple
independent class estimations.
During the past decade, a number of initiatives have been undertaken to create systematic
national and global data sets of processed satellite imagery. An important application of these data
is the derivation of large geographic area (i.e., multiscene) LC products. These products exhibit
internal variations in information quality for two principal reasons. First, they have been assembled
from a multitemporal mix of satellite scenes acquired under differing seasonal and atmospheric
conditions. Second, intraproduct landscape diversity will lead to spatially varying levels of class
commission errors. Detailed modeling of these variations with conventional ground truth is pro-
hibitively expensive, and hence an alternative accuracy assessment method must be sought.
In this chapter we presented a method for confidence estimation based on the analysis of
classification consistency in regions of overlapping coverage between Landsat scenes from adjacent
orbital paths and rows. A LC mapping methodology has been developed that exploits consistency
evaluation to (1) improve scene-based classification performance, (2) support the integration of
scene classifications through compositing, (3) provide a detailed confidence characterization of the
final product, and (4) conduct postgeneration accuracy assessment. This methodology was imple-
mented within a prototype mapping system, QUAD-LACC, to derive synoptic LC products of the
Laurentian Great Lakes watershed. It should be noted that others researchers have suggested using
overlap regions to assess the accuracy of landscape metrics (Brown et al., 2000).
Brown, D.G., J.D. Duh, and S.A. Drzysga, Estimating error in the analysis of forest fragmentation change
using North American landscape characterization (NALC) data,
Remote Sens. Environ.
, 71, 106-117,
Congalton, R.G., A review of assessing the accuracy of classifications of remotely sensed data,
, 37, 35-46, 1991.