To assess the accuracy of the classification results
from the decision trees, we first conducted field
sampling and then used the ground truth to build
error matrices. The field sampling adopted a
stratified random sampling so that the distribution
of the sample points was relatively even across
different land cover categories (Congalton 1991).
Besides the correctness for each category that can be
directly read from the error matrix (Hord and
Brooner 1976, Congalton 1991), we also calculated
more sophisticated Kappa coefficients to measure
differences between actual and chance agreement
(Cohen 1960, Fung and LeDrew 1988, Congalton
1991, Campbell 2002). The error matrices for the
three years and their corresponding Kappa values
are listed in Table 6.
It was not surprising that results of 2002 had the
highest accuracy because theTMimage from this year
was used to build the decision tree. However, we
consider the results from the other two years to also be
satisfactory. Their relatively high accuracies indicate
that a decision tree built using the data from one year
may be reused for other years in identifying mangrove
forests. As a result, we might be able to establish a
general classification rule set for this application.