3.1 Land use/cover classification and accuracy
assessment
Our classified land-cover maps for 1985, 1998, and 2016 consisted of
‘rocks’, ‘rangeland type I’, ‘rangeland type II’, ‘water’,
‘agriculture’, ‘residences’, ‘forest’, ‘industry’ and ‘other’ thematic
classes, and were produced with plausible classification accuracies
(figure 2, Table 2). To increase the accuracy for the ‘agriculture’
class, lands with clean and green fallow, tilled soil and lands under
crops were combined. To increase the accuracy of the ‘residences’ class,
post-classification assessment was performed with a reverse comparison
of the residences’ patterns; if they first appeared in 1985, they should
also appear in the 1998 and 2016 maps (the assumption being that there
was no reduction in residential area over time).
<Figure 2>
Table 2 shows the accuracy of maps and classes produced with a random
forest classifier. The ‘agriculture’ thematic class had a high accuracy
in all years. The accuracies for the ‘rangeland type I’, and ‘forest’
thematic classes were high as well. The user’s and producer’s accuracies
for the ‘residences’ class varied from 0.67 to 0.80. The classes
‘rocks’, ‘other’, and ‘rangeland type I’ were the least accurate ones
due to low spectral separability among these classes.
<Table 2>