A backcast land use change model to generate past land use maps: application and validation at the Muskegon River watershed of Michigan, USA

We developed a GIS and neural network-based land use/land cover change model for backcasting land use change and applied it to the Muskegon River watershed, a typical upper midwestern watershed in the USA. We developed 12 variants of the model, based on different structural assumptions, to simulate urban, forest, agriculture, and shrubland transitions. We compared the model variants against 12,598 land use interpreted locations from 235 aerial photographs acquired from the study region between the late 1930s through to the early 1970s. The model variants produced around 41-70% accuracy (integrating both omission and commission errors) in simulating the spatial locations of the dominant land use category, forests and agriculture, and lower accuracy for the shrub and urban land use categories. We describe the assumptions made in developing the model and discuss the implications of the assumptions to model goodness-of-fit analysis and to forecasting land use.