A novel strategy to assimilate category variables in land-use models based on Dirichlet distribution

Data assimilation is an effective approach to reduce the propagation and cumulative errors of land use models. However, as the discrete categorical outputs of land use models, land use data assimilation requires a novel approach distinguished from the traditional assimilation for continuous variables. Here, Bayesian inference for categorical distribution is introduced into a land use cellular automata model to update multiple discrete state variables. The accuracies with data assimilation outperform those of simulation-only. By 2009, kappa coefficient and figure of merit in the entire area increase by 0.34% and 1.78%, respectively, and in fine-scale areas where drastic and representative land use changes occurred, increase by 23.88% and 38.39%, respectively. The assimilation performance is associated with the landscape patch size and the length of the assimilation cycle. This study innovatively introduces the conjugate prior features of Dirichlet distribution into land use assimilation, providing insights and references for discrete variable data assimilation.

Automated summary

Data assimilation is an effective method for reducing errors in land use models. This study introduces Bayesian inference to update discrete state variables in a land use cellular automata model, improving assimilation accuracy. The performance of assimilation is influenced by patch size and assimilation cycle length.