The modelled land cover was then compared to the observed land cover using the Kappa statistic (Congalton, 1991; Jensen, 1996). Two successivemajority filters using a 3 3 neighbourhood provided the best predictions of reality (37% total accuracy) and were therefore used in the SIENA land over model. Minor roads were modelled for a 250 250 m grid because sensitivity analysis (results not shown) established this as the highest possible resolution which allowed sufficient minor road density variation between the grid cells to inform the probabilistic model. As outlined in Table 3, the minor road density was grouped into fifteen density classes. This provided a wide distribution of densities with approximately equal numbers of grid cells in each class. Again, class membership was constrained using a stress measure to avoid over-representation of some density classes. The population model reflected the fact that each land cover class had a characteristic population density. First, population was calculated at land cover patch level by applying the mean population density of the inland cities for each land cover class to the SIENA land cover patches as follows: