Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 150

calculated by dividing the number of correct samples of (c) by the total number of
reference samples of class (c). The resulting percentage producer accuracy indi-
cates the probability that a reference pixel will be correctly classified.
Producer accuracy ¼ n cc
n þ c
The user accuracy, that is a measure of error of commission (Story and
Congalton 1986 ), describes how many samples that were classified as (c) in fact
belong to class (c). The measurement is resulted from:
User accuracy ¼ n cc
n c þ
Finally, multiplying the results of each the previous three accuracies by 100
forms the percent correctly classified (PCC) metric.
The second statistic used is the kappa coefficient (kc). It is generally known as a
precision measure since it is considered as a measure of agreement in the absence
of chance (Cohen 1960 ; Lillesand et al. 2008 ). Conceptually it can be defined as:

KC ¼ Observed Accuracy Chance Agreement
1 Chance Agreement

The kappa statistic is calculated from the confusion matrix by using the fol-
lowing mathematical statement:
KC ¼ n P i ¼ 1 x ii P i ¼ 1 x io x oi

n 2 P i ¼ 1 x io x oi

where:
n total number of pixels used for testing the accuracy of a classifier,

p number of classes,

P x ii

sum of diagonal elements of confusion matrix,

P x io

sum of row i,

P x oi

sum of column i

An example is presented to explain the derived classification accuracies for the
21,000 ha project for the year 2007 (Table 5.9 ).
The second method is based on the state administrative divisions (e.g., Menbij
see Sects. 5.4 and 5.7.2 ) and/or on the state irrigation projects divisions (e.g., the
21,000 ha project see Sect. 5.10 ), to derive the accuracy of the correspondence
between the derived statistical numbers from the automated classification of
remote sensing data and those human-based statistical records.
The correspondence degree for a specific season at an administrative-scale was
measured by calculating the Percent Error (PE):

PE ¼ Observed i Predicted i
Observed i

100 :