Modelling land cover transitions: A solution to the problem of spatial dependence in data

Abstract

Raster-based spatial land cover transition models (LCTMs) are widely used in landscape ecology. However, many LCTMs do not account for spatial dependence of the input data, which may artificially fragment the output spatial configuration. We demonstrate the consequences of ignoring spatial dependence, thus assigning probabilities randomly in space, using a simple LCTM. We ran the model from four different initial conditions with distinct spatial configurations and results indicated that, after 20 simulation steps, all of them converged towards the spatial configuration of the random data set. From an ecological perspective this is a serious problem because ecological data often exhibit distinct spatial configuration related to ecological processes. As a solution, we propose an approach (region approach) that accounts for spatial dependence of LCTM input data. Underlying spatial dependence was used to apply spatial bias to probability assignment within the model. As a case study we applied a region approach to a Vegetation Transition Model (VTM); a semi-Markovian model that simulates forest succession. The VTM was applied to approximately 500,000 ha of boreal forest in Ontario, at 1 ha pixel resolution. When the stochastic transition algorithms were applied without accounting for spatial dependence, spatial configuration of the output data became progressively more fragmented. When the VTM was applied using the region approach to account for spatial dependence output fragmentation was reduced. Accounting for spatial dependence in transition models will create more reliable output for analyzing spatial patterns and relating those patterns to ecological processes.