Markov random field models for vector-based representations of landscapes

Abstract

In agricultural landscapes the spatial distribution of cultivated and seminatural elements strongly impacts habitat connectivity and species dynamics. To allow for landscape structural analysis and scenario generation, we here develop statistical tools for real landscapes composed of geometric elements, including 2D patches but also 1D linear elements (e.g., hedges). Utilizing the framework of discrete Markov random fields, we design generative stochastic models that combine a multiplex network representation, based on spatial adjacency, with Gibbs energy terms to capture the distribution of landscape descriptors for land-use categories. We implement simulation of agricultural scenarios with parameter-controlled spatial and temporal patterns (e.g., geometry, connectivity, crop rotation), and we demonstrate through simulation that pseudo-likelihood estimation of parameters works well. To study statistical relevance of model components in real landscapes, we discuss model selection and validation, including cross-validated prediction scores. Model validation with a view toward ecologically relevant landscape summaries is achieved by comparing observed and simulated summaries (network metrics but also metrics and appropriately defined variograms using a raster discretization). Models fitted to subregions of the Lower Durance Valley (France) indicate strong deviation from random allocation and realistically capture landscape patterns. In summary, our approach improves the understanding of agroecosystems and enables simulation-based theoretical analysis of how landscape patterns shape biological and ecological processes.