A hybrid analytical-heuristic method for calibrating land-use change models

Spatially explicit land-use models simulate the patterns of change on the landscape in response to coupled human–ecological dynamics. As these models become more complex involving larger than ever data sets, the need to improve calibration techniques as well as methods that test model accuracy also increases. To this end, we developed a Genetic Algorithm tool and applied it to optimize probability maps of deforestation generated from the Weights of Evidence method for 12 case-study sites in the Brazilian Amazon. We show that the Genetic Algorithm tool, after being constrained during the reproduction process within a specified range and trend of variation of the Weights of Evidence coefficients, was able to overcome overfitting and improve validation fitness scores with acceptable computational costs. In addition to modeling deforestation, the Genetic Algorithm tool coupled with the Weights of Evidence method is flexible enough to embrace a variety of models as well as their specific fitness functions, thus offering a practical way to optimize the performance of land-use change models.

► A hybrid calibration method allows the optimization of transition probability maps.
► The hybrid method takes advantage of Weights of Evidence and Genetic Algorithm.
► We evaluated several map comparison methods to choose the best fitness function.
► The hybrid approach overcame overfitting and produced more accurate predictions.
► The GA tool offers a practical way to optimize a variety of environmental models.