Modeling deforestation baselines using GEOMOD - Instituto ...

Finalizing Avoided Deforestation Baselines
Driver 4
Driver 3
Driver 2
Driver 1
Calibration
Procedure
Likelihood
map
T1 Landuse
Map
cells deforested = y
Simulated T2
Map
cells deforested = y
T2 Landuse
Map
Validation
Procedure
Kappa for
Location
Figure 1. Flowchart for GEOMOD showing how empirical knowledge (maps) is used to weight
drivers and how, through comparison of simulated results versus the actual landscape, we use the
Kappa-for-location measurement of ‘goodness of fit’ to validate which driver set explains most of
the spatial variation of land use change.
III. C. Validation
To validate the results created by GEOMOD, the actual land-use map at a known point in time is compared
to the simulated land-use map of that same time period based on analysis and projection of the pattern of
land use from the earlier point in time. It was common in the past to measure ‘goodness of fit’ using a
simple percent correct measure or, at best, a multiple-resolution percent correct (Costanza, 1989, Hall et al.
1995) but this provides little assessment of a model’s ability to predict the correct quantity of change versus
its ability to identify the correct location of change. Spatial measures of ‘goodness of fit’ have been
developed that measure the degree to which a simulated map agrees with a reality map with respect to
both location (Kappa-for-location) and quantity of cells correct (Kappa-for-quantity) (Pontius, 2000).
How then do we explicitly measure success? First we determine how many cells did GEOMOD get right.
This is the model’s success rate. Then, what is the expected success due to chance alone? This we call
the success rate based on random cell selection and is calculated as:
R = (C2 + (N – C) 2) / N
where:
R = the rate of success expected from random cell selection
C = number of grid cells that change
Winrock International
A- 4