Page 796 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

(a) (b)

Fig 3. (a) natural logarithm of road, (b) square root of river

The distance from road parameter was considered dynamic and expandable in the time series, so that it was first classified
st
nd
into 3 levels, includiing: basic roads, 1 grade roads, and 2 grade roads. Historical behaviour between 2002 and 2008 and the
st
earth's digital elevation model were used as training dataset. Then, the 2 grade side roads have become the 1 grade side roads
st
st
and the 1 grade side roads have become the major roads, and the basic ways were expandable at the beginning and the end using
two conditions: a) Minimum gradient in the direction of the slope of the area, and b) Maximum change potential.
Firstly, the map of distance from road for the target year was predicted with this method, then it was used as an
independent factor in predicting the target year map.

4. Proposed Method

4.1. Similarity weighted instance‐based learning

Weight-based learning method is one of the transfer potential modelling methods that follows the logic of k-nearest
neighbour method. This method is based on the calculation of weight distances to variable space for known samples of land use
classes. Basically, in connection with the creation of transfer potentials to model the change in the land cover, the model considers
two classes for each transfer (fixed pixels and variable pixels) and evaluate them. In this method, the k-nearest neighbour for
each fixed or variable pixel (pixel with a position that it’s changes is considered for the future) is examined, and then the distances
in the variable space from each unknown pixel to the pixels around (range k), which they have changed during the calibration
period are calculated[7][1] (see Fig. 4).
This distance is obtained in an exponential function in order to calculate a continuous level of membership of the existing
land use layers (classes) for each pixel from the following equation[3]:

1

∑ =1 (1 − 1 ) (1)

ℎ ℎ = 1 +

where c is the number of variable pixels in the range k, d is a linear distance to variable pixels in the range k, and K is the total
number of variable and constant pixels in the range k.
There are several points to note about the learning process based on the similarity weight model, which is mentioned
below:
• The difference between this method and the method of k-nearest neighbour is that this method considers the
continuous level of the membership class instead of considering the strict value.

5