Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 148

Classification
results/Thematic maps
to be assessed

Reference truth data sets
(GPS-points, aerial photographs, etc.)

Accuracy evaluation
(error/confusion matrix)

Create random points

Classification accuracy
evaluation (producer
accuracy, user
accuracy, and overall
accuracy)

Fig. 5.62 The general accuracy assessment steps that were applied on the resulted thematic map/s
from the classification process

be 75-100 samples for each LULC-category. This suggested approach samples
small areas thoroughly, while large areas might be under-sampled. Thus, it is
suggested that testing sample numbers could be set for variations in size and
within-class variability.
Accuracy assessment is a post-classification step. It was accomplished for the
purposes of this study using ENVI, v. 4.6, which was used to evaluate the cor-
respondence of the classified LULC-maps to the true and/or assumed true geo-
graphical reference data (Congalton 1991 ). The reference data were: Part of the
collected field-data for the years 1987, 2005 and 2007(see Sect. 5.4 ), where the
first part was used as training samples; assumed truth data based on the integration
of the remotely sensed data; irrigation projects statistical records and the detailed
construction schemes of these projects, which were used locate the spatial distri-
bution of the various agricultural features in the irrigation projects area for 1987
(see Sect. 5.10 ); thematic maps; visual interpretation based on the remote sensing
data itself; and Google Earth. Figure 5.62 illustrate the major steps that were
followed in assessing the various thematic maps that resulted from the classifi-
cation process.
Results of classification were presented in form of thematic maps. Using the
various truth reference data, accuracy assessments were carried out for all clas-
sification results. The reference data/classes were compared with the predicted
classes by the adopted classifier/s (and probably enhanced using the post-classi-
fication processing). The final evaluation results were reported in the form of error
matrices. The overall classification accuracy (percentage correct) was calculated
for all classifications, as well as the accuracies of the class-specific user and
producer.
Two accuracy assessment methods were performed in this thesis. The first
method is based on the pixel scale to derive the accuracy of classification in the
remotely sensed data, which resulted from the calculation of the error/confusion
matrix.