Impact and Integration of Multiple Training Dates for Markov Based Land Change Modeling

Abstract

Most geomatic land use/cover (LUC) simulation tools use two LUC maps as training dates, particularly prediction models based on Markov chains. In this paper we begin by listing the potential errors resulting from only considering two past dates. We then illustrate the consequences of this approach on quantitative model calibration using a dataset encompassing six LUC maps. This offers multiple Markovian combinations for input matrices generating a wide range of Markovian probability transitions. An even larger spectrum can be achieved by introducing limited confidence in data. The comparison of LUCC budgets and possible Markov chains offers a broad spectrum of results and randomness in the choice of only two dates. We propose two techniques for integrating the knowledge obtained from more than two training dates into forecasting scenarios. First we calculate an annual rate of change, which is weighted according to time distance from the present in order to fix expected total change in the simulation step and at the category level. We then produce alternatives to Markov chains at a transitional level. In this way we integrate all available LUCC-budgets and propose different methods for weighting observed transitions, so as to produce transition matrices that could act as alternatives to Markov chains based on just two dates.