Model 1 23x25x1 Output, Target
and Errors
Map
of the predicted European population surface
Close
up of predictions for Great Britain
Surpop Target
data
Error
map at 1DM resolution for Great Britain (district constraints)
SAS ward
population count
Ward
population prediction (district constraints)
Difference
between SAS ward population and prediction
Map
of predictions for England and Wales from ward constraints
Error
map at 1DM resolution for England and Wales (ward constraints)
Close
up of predictions for Italy
Registration
zone population
Registration
zone population prediction (Nuts3 constraints)
Difference
between Registration zone population and prediction
Map
of predictions for Italy from Registration data zone constraints
Comments
-
The 23x25x1 neural network interpolates from Nuts3 level to Nuts5 level
fairly well.
-
The largest absolute errors occur where population density is highest in
urban areas.
-
Calculating the predicted Ward and Registration zone populations involved
summing the predictions for all 1DM cells which lie in each Ward or Registration
zone polygon. In order to do this the grid of predictions was integerised
and converted into a polygon coverage in ArcInfo and the population density
of each polygon region was calculated. The polygon coverage was intersected
with the Ward or Registration polygons and the predicted population of
each small intersected area was calculated by multiplying by the predicted
population density. Finally the predicted population in each Ward or Registration
zone was found by summing the population of small intersected areas for
each Ward or Registration zone polygon.
-
The error in Great Britain is greatly reduced using Ward level constraints.
A pattern in the errors appears which could be something to do with differences
between SAS and surpop populations and maybe worth investigating further.
-
The Italian surface also appears greatly improved using the Registration
zone constraints, but there is no real evidence for this.
-
The main sources of error include; census data error, those resulting from
imperfections in neural network prediction rescaling, and those relating
to an inappropriately small selection of training data cells with high
population density.