Identifying dominant factors for the calibration of a land-use cellular automata model using Rough Set Theory

While cellular automata (CA) are increasingly used for modeling urban growth and land-use changes, the methods for identifying the dominant factors that drive the landscape dynamics when calibrating the model still require improvement, specifically in the context where a large number of factors are considered. In this paper, the potential of Rough Set Theory (RST) to guide the factor selection is evaluated. This data mining approach was tested for the calibration of a CA model to simulate land-use changes in a portion of the Elbow River watershed adjacent to the City of Calgary, in southern Alberta, Canada. Simulation outcomes obtained using a total of 18 original factors and a smaller set of factors identified with RST were compared to reference land-use maps using three Kappa coefficients of agreement. Results reveal that the factors selected by RST are not identical for each land use. Among the identified factors, three external factors (distance to river, distance to Calgary City center, and distance to road) and the presence of Built-up areas in the three considered neighborhoods are the most important factors driving the transition from Forest and Vegetation (including agriculture and Rangeland/Parkland) to Built-up. The Kappa statistics reveal that the factors selected by RST tend to generate a higher agreement with reference land-use maps than the original group of 18 factors and that they are better at capturing quantity information than location information. An advantage of RST is that it retains the original factors in the identification of the transition rules. In addition, the computation time required for the simulation using the RST factors is considerably less than the time needed to generate the results using the original set of factors. However, the data mining technique itself is computationally intensive. This study illustrates that RST can guide the selection of the dominant factors required in the calibration of a CA model, but that its potential still needs to be further investigated.

Research highlights
► The goal of the study is to evaluate the potential of a data mining technique, Rough Set Theory (RST) to guide the selection of driving factors for the calibration of a land-use cellular automata (CA) model.
► The factors selected by RST are not identical for each land-use transition simulated by the model.
► The reduced number of factors selected by RST tends to generate a higher agreement with reference land-use maps than the original set of factors used for the calibration of the CA model.
► An advantage of RST is that it retains the original factors in the identification of the transition rules.
► The computation time required for the simulation using the RST factors is considerably reduced; however the selection of the factors itself is computationally intensive.