Spatial scale in land use models: Application to the Teruti-Lucas survey

We consider the problem of land use prediction at different spatial scales using point level data such as the Teruti-Lucas (T-L hereafter1) survey and some explanatory variables. We analyze the components of the prediction error using a synthetic data set constructed from the Teruti-Lucas points in the Midi-Pyrénées region and a five categories land use classification. The study first shows that the number of points in the Teruti-Lucas survey is quite enough for estimating the probabilities of each land use category with a good quality. Furthermore it reveals that, contrary to usual practice, when the objective is to predict land use at aggregated levels, land use probabilities should be estimated at more locations where explanatory variables are available rather than restricting to the initial Teruti-Lucas locations. Indeed this strategy borrows strength from the knowledge of the explanatory variables which may be heterogeneous. Finally, guidelines for constructing the grid of locations for estimation are given from the analysis of the heterogeneity of each explanatory variable.