ap 
1g 
1g 
ial 
ata 
Ire 
Ily 
the 
ial 
ind 
irst 
n a 
ith 
‚of 
  
  
Felkner, John 
2 Land Use Change Prediction: A Statistical Tree Classification Approach 
For both the independent factor maps of land use change 1979-1989 and the input environmental and socioeconomic 
input factors, the value of each grid cell or pixel in each digital map is fed into the statistical software and a tree 
classification is performed (such an operation is referred to as a tree "classification" if the dependent variable is 
categorical, and a tree "regression" if the dependent variable is continuous). The tree classification process also allows 
for the prediction of the probabilities of each observation converting to a particular change class. For example, for 
pixel x,y, the tree classification will produce probabilities of that pixel converting to one of the four land use change 
categories: conversion of forest to agriculture; conversion of agriculture to urban; conversion of forest to urban; and no 
change. 
The tree classification algorithm works by performing a binary split on the data that maximizes the overall reduction in 
deviance. The measurement of deviance at any particular “node” of the tree (where a split is performed) is measured by 
D: = -2 Nix log Di ( 1) 
k 
at each node i of the tree, with the probability distribution p over the classes as that node and the random sample of 
observations n at that node i. The deviance value for the entire tree is then given by 
D- Di Q) 
for all nodes (Venables and Ripley 1994). 
Since the tree classification has as the dependent variable observations that reflect land use change from 1979 to 1989, 
and since the independent variables are values that depict general conditions for selected economic and environmental 
variables during the 1980s, the tree classification predicted probabilities can be used to create a predicted map of land 
use change for 1999. 
These probabilities of change for each pixel can be fed back into the GIS, and maps of probabilities can be created. For 
example, a map can be created in which each pixel in the map represents the probability of that pixel converting to, say, 
urban use in the future. Another map can be created showing the probability of each pixel converting from forest to 
agriculture, etc. These maps can be combined into one map through a series of overlay operations, assuming that 
priority is given for conversion of either forest or agriculture to urban. This single map can be said to represent 
probabilities of future land use conversion. 
In order to finish the prediction process, a map of landcover from 1999 was derived from 1999 Landsat 7 satellite 
imagery. This map is compared with the 1989 landcover images. This comparison allows calculation of the number of 
pixels of new urban, new agriculture, etc., that actually occur from 1989 to 1999. These numbers of new pixels — the 
“demand” quantities for land use change — are then allocated on the predicted probability maps in rank order, 
prioritizing urban change first, then agriculture, etc. 
24 Comparison Between Actual and Predicted Land Use Change 
Comparison between the predicted and actual landcover maps is performed using systematic GIS comparison methods : 
that allow for measures of prediction accuracy. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 435