Calibrating a neural network-based urban change model for two metropolitan areas of the Upper Midwest of the United States

We parameterized neural net-based models for the Detroit and Twin Cities metropolitan areas in the US and attempted to test whether they were transferable across both metropolitan areas. Three different types of models were developed. First, we trained and tested the neural nets within each region and compared them against observed change. Second, we used the training weights from one area and applied them to the other. Third, we selected a small subset (∼1%) of the Twin Cities area where a lot of urban change occurred. Four model performance metrics are reported: (1) Kappa; (2) the scale which correct and paired omission/commission errors exceed 50%; (3) landscape pattern metrics; and (4) percentage of cells in agreement between model simulations. We found that the neural net model in most cases performed well on pattern but not location using Kappa. The model performed well only in one case where the neural net weights from one area were used to simulate the other. We suggest that landscape metrics are good to judge model performance of land use change models but that Kappa might not be reliable for situations where a small percentage of urban areas change.