2.3 Classification with random forest classifier
Image classification is a method for extracting environmental
parameters, such as land cover and land use (LCLU) (Jensen, 2004). An
expert who conducts a supervised classification should have prior
knowledge about study area conditions and land uses. For this study,
training samples were determined by using available maps, aerial images,
field data and ancillary spatial information, such as soil types. We
used a random forest classifier implemented in the package “random
Forest” in statistical R software. Based on the literature and
preliminary tests, we used 1,000 trees to parameterize the random forest
classifier. We used the Blue, Green, Red, NIR, SWIR-1, and SWIR-2 bands
for Landsat 4, 5 TM, Landsat 7 ETM+ and Landsat 8 OLI to produce LCLU
change maps with a random forest classifier. The iterative
classification approach was used with an increase of training samples
selected for each thematic class until the classification accuracy for
individual thematic classes did not increase anymore.
We developed a classification catalog, which presented a mix of LCLU
classes (Table 1). Our classification catalog consisted of eight
thematic classes: ‘forest’, ‘agriculture’, ‘rangeland’, ‘water’,
‘rocks’, ‘industry’, ‘residences’, and ‘other’. The ‘forest’ class
consisted of urban and industrial green space with trees and tree
plantations, ‘agriculture’ consisted of tilled agricultural land used
for crops and clean or green fallow land as a part of crop rotation. The
‘rangeland’ class consisted of native natural plants with low (type I)
to medium (type II) density. ‘Water’ represented water objects, such as
rivers, natural wetlands and man-made water ponds. The ‘rock ‘class
consisted of rocky outcrops and the mountains. ‘Industry’ consisted of
industrial areas and mines. ‘residences’ comprised urban and rural
settlements and main paved roads and the ‘other’ class was bare soil,
industrial and mineral wastelands, and dirt roads.
<Table 1>