Classification of Urban Areas: Inferring Land Use from the Interpretation of Land Cover - Remote Sensing of Urban and Suburban Areas - page 144

Fig. 8.3 Maximum likelihood classification results using unequal prior probabilities for the city of
Bristol. Note the differences in the three housing types between unequal priors in the three high-
lighted census tracts and results from equal priors shown in Fig. 8.2 . In all these cases the equal prior
probability classification has over-predicted (or over classified) detached housing (low density) and
under predicted semi-detached and terraced (medium and low density). This is most apparent for
09DDFZ29 where no detached housing exists. Under unequal prior conditions, detached and semi-
detached are also over-predicted but this time terraced is very close to census estimates
prior knowledge has resulted in the misclassification of detached dwellings. In the
case of the census tract (09DDFZ29) within the city center detached housing is
again classified using equal priors despite the fact that none are recorded in the
census. This zero prior probability is included within the stratified and modified
method and results in no pixels classified as detached. What the results reveal is that
total absolute error is lower under modified prior probabilities, but that no one
housing type has consistently lower area estimation error. However, there is some