Elsevier

Spatial Statistics

Volume 6, November 2013, Pages 118-138
Spatial Statistics

A completely random T-tessellation model and Gibbsian extensions

https://doi.org/10.1016/j.spasta.2013.09.003Get rights and content

Abstract

In their 1993 paper, Arak, Clifford and Surgailis discussed a new model of random planar graph. As a particular case, that model yields tessellations with only T-vertices (T-tessellations). Using a similar approach involving Poisson lines, a new model of random T-tessellations is proposed. Campbell measures, Papangelou kernels and formulae of Georgii–Nguyen–Zessin type are translated from point process theory to random T-tessellations. It is shown that the new model shows properties similar to the Poisson point process and can therefore be considered as a completely random T-tessellation. Gibbs variants are introduced leading to models of random T-tessellations where selected features are controlled. Gibbs random T-tessellations are expected to better represent observed tessellations. As numerical experiments are a key tool for investigating Gibbs models, we derive a simulation algorithm of the Metropolis–Hastings–Green family.

Introduction

Random tessellations are attractive mathematical objects, from both theoretical and practical points of view. The study of the mathematical properties of these objects is still leading to open problems, while the range of applications covers a broad panel of scientific domains such as astronomy, geophysics, image processing or environmental sciences (Lantuéjoul, 2002, Møller and Stoyan, 2007, Le Ber et al., 2009).

When modeling real-world structures, one aims at flexible classes of random models able to represent a wide range of spatial patterns. Gibbsian point processes combined with Voronoï diagrams offer such an attractive approach. The class of Gibbs point processes is enriched continuously (hard-core, Strauss, area-interaction, Quermass-interaction). Random tessellations are for example obtained as Voronoï diagrams of seeds distributed according to such Gibbs processes, see e.g. Dereudre and Lavancier (2011). Of special interest are Gibbs models with interactions based on the Delaunay graph (Baddeley and Møller, 1989, Bertin et al., 1999, Dereudre et al., 2012): Delaunay-neighbor seeds define Voronoi cells with a common edge. Hence a large class of models for random tessellations is made available for applications. Our aim is to sketch a similar theoretical framework for other types of tessellations which are not seed-based. We will focus on T-tessellations: tessellations with only T-vertices.

An example of a stochastic model for T-tessellation is the STIT (STable with respect to ITeration) tessellation model (Nagel and Weiss, 2005). STIT tessellations are obtained by successive splits and rescaling. Analytical results about the distributions of various geometrical features are available (e.g.  Mecke et al., 2007, Thäle, 2011, Cowan, 2013, Schreiber and Thäle, 2010, Weiss et al., 2010). Ergodicity and mixing properties of the STIT model have also been investigated (Martínez and Nagel, 2012, Lachièze-Rey, 2011). Recently generalizations of the STIT model involving different splitting procedures have been considered (Cowan, 2010, Schreiber and Thäle, 2013). Such random tessellations are referred to as nested tessellations following the denomination used in Schreiber and Thäle (2013). Realizations of nested-tessellations show a striking feature: any compact convex region is split by a unique tessellation maximal segment (also called I-segment).

Another example of stochastic model is the Gilbert model (Gilbert, 1967, Mackisack and Miles, 1996) based on segments growing until they are blocked by other segments. Analytical results on the distributions of standard geometrical features are more sparse (Mackisack and Miles, 1996, Burridge et al., 2013). Some asymptotic results have been established (Schreiber and Soja, 2011). There are also some restrictions on the geometry of tessellations arising from the Gilbert model. Since segments are born simultaneously and grow at a common speed, the number of other segments a segment can block is limited. Therefore, again, arbitrary T-tessellations cannot be obtained by a Gilbert-type construction.

The models discussed in this paper define random T-tessellations with realizations from a large class of T-tessellations built using three geometrical operators: splits, merges and flips. Our approach is closely related to the one used by Arak, Clifford and Surgailis in their paper (Arak et al., 1993) on a random planar graph model. In particular, a key ingredient is the Poisson line process. As a first step, this paper introduces a new model of random T-tessellations which can be considered as a completely random model. Then Gibbs variations are proposed and a general algorithm for simulating them is described.

Throughout the paper, we focus on the case where the domain of interest is bounded. Extension of Gibbs models for T-tessellations of the whole plane remains an open problem at this stage.

Section  2 provides definitions, notation and basic results about T-tessellations. The completely random T-tessellation model is discussed in Section  3. This section also introduces for arbitrary random T-tessellations Campbell measures, Papangelou kernels which are widely used in point process theory. Our new model can be considered as a T-tessellation analogous to the Poisson point process. This claim is based on Georgii–Nguyen–Zessin type formulae. Therefore the T-tessellation model introduced in Section  3 is referred to as a completely random T-tessellation. Gibbsian extensions are discussed in Section  4 together with some examples. One example is also a particular case of Arak–Clifford–Surgailis random graph model when its parameters are chosen in order to yield a T-tessellation. Formulae of Georgii–Nguyen–Zessin type are provided for hereditary Gibbs models. In Section  5, a simulation algorithm is proposed. The design of the algorithm follows the general principles of Metropolis–Hastings–Green algorithms, already widely used for Gibbs point processes. It involves three types of local operators: split, merge and flip. Conditions ensuring the convergence of the Markov chain to the target distribution are provided.

Below, as a notational convention, bold letters are used for denoting random variables. Detailed proofs of the main results are postponed in appendices.

Section snippets

The space of T-tessellations

In this paper, we shall consider only tessellations of a compact domain D in R2. For the sake of simplicity, D is supposed to be also convex and polygonal. Let a(D),l(D),ne(D) and nv(D) be respectively the area, the perimeter length, the numbers of edges and vertices of D.

A polygonal tessellation of D is a finite subdivision of D into polygonal sets (called cells) with disjoint interiors. The tessellation vertices are cell vertices. Edges are defined as line segments contained in cell sides,

A completely random T-tessellation

Let us start with a formal definition of a random T-tessellation. From now on, a T-tessellation is considered as a closed set defined as the union of its edges (or segments). The space T is equipped with the standard hitting σ-algebra σ(T) (see Matheron, 1975) generated by events of the form {TT:(eE(T)e)K} where K runs through the set of compact subsets of D. A random T-tessellation is a random variable T taking values in (T,σ(T)).

Our candidate of completely random T-tessellation is

Gibbsian T-tessellations

Although the completely random T-tessellation model introduced in the previous section shows appealing features, it may not be appropriate for representing real life structures which may exhibit some kind of regularity. This section is devoted to Gibbsian extensions allowing to control a large spectrum of T-tessellation features. Gibbs random T-tessellations are defined as follows.

Definition 5

Let h be a stable non-negative functional on T. The Gibbs random T-tessellation with unnormalized density h is the

A Metropolis–Hastings–Green simulation algorithm

In this section, we derive a simulation algorithm for random Gibbs T-tessellations. This algorithm is a special case of the ubiquitous Metropolis–Hastings–Green algorithm, see e.g. Geyer and Møller (1994), Green (1995) and Geyer (1999). It consists of designing a Markov chain with state space T and with invariant distribution the target probability measure P.

The design of a Metropolis–Hastings–Green algorithm involves two basic ingredients: random proposals of updates and rules for accepting or

Discussion

The main feature of the completely random T-tessellation introduced in this paper is that both split and flip Papangelou kernels have very simple expressions showing a kind of lack of spatial dependency. It would be of great interest to further investigate this model. Since analytical probabilistic results are available for the Arak–Clifford–Surgailis (Example 2), it is expected that such results could also be derived for our model. In particular, the following issues are of interest:

  • Is there a

Acknowledgements

The authors would like to thank the anonymous referees for careful reading and helpful comments. The authors also wish to thank their colleagues who showed interest in this work presented at several workshops since 2006.

References (40)

  • D. Dereudre et al.

    Practical simulation and estimation for Gibbs–Delaunay–Voronoi tessellations with geometric hardcore interaction

    Comput. Statist. Data Anal.

    (2011)
  • F. Le Ber et al.

    Neutral modelling of agricultural landscapes by tessellation methods–application for gene flow simulation

    Ecol. Modell.

    (2009)
  • T. Arak et al.

    Point-based polygonal models for random graphs

    Adv. Appl. Probab.

    (1993)
  • T. Arak et al.

    Markov fields with polygonal realizations

    Probab. Theory Related Fields

    (1989)
  • A. Baddeley et al.

    Nearest-neighbour Markov point processes and random sets

    Int. Stat. Rev.

    (1989)
  • E. Bertin et al.

    Existence of Delaunay pairwise Gibbs point process with superstable component

    J. Stat. Phys.

    (1999)
  • J.-M. Billiot et al.

    Maximum pseudolikelihood estimator for exponential family models of marked Gibbs point processes

    Electron. J. Stat.

    (2008)
  • J. Burridge et al.

    Full- and half-Gilbert tessellations with rectangular cells

    Adv. Appl. Probab.

    (2013)
  • R. Cowan

    New classes of random tesselllations arising from iterative division of cells

    Adv. Appl. Probab.

    (2010)
  • R. Cowan

    Line segments in the isotropic planar STIT tessellation

    Adv. Appl. Probab.

    (2013)
  • D. Dereudre et al.

    Existence of Gibbsian point process with geometry-dependent interactions

    Probab. Theory Related Fields

    (2012)
  • C. Geyer

    Likelihood inference for spatial point processes

  • C.J. Geyer et al.

    Simulation procedures and likelihood inference for spatial point processes

    Scand. J. Stat. Theory Appl.

    (1994)
  • E.N. Gilbert

    Random plane network and needle-shaped crystals

  • P.J. Green

    Reversible jump Markov chain Monte Carlo computation and Baysian model determination

    Biometrika

    (1995)
  • J.L. Jensen et al.

    Pseudolikelihood for exponential family models of spatial point processes

    Ann. Appl. Probab.

    (1991)
  • Kahn, J., 2010. Existence of Gibbs measure for a model of T-tessellation. December....
  • O. Kallenberg

    Random Measures

    (1986)
  • R. Kluszczyński et al.

    Image segmentation by polygonal Markov fields

    Ann. Inst. Statist. Math.

    (2007)
  • R. Lachièze-Rey

    Mixing properties for STIT tessellations

    Adv. Appl. Probab.

    (2011)
  • Cited by (18)

    • How can models foster the transition towards future agricultural landscapes?

      2021, Advances in Ecological Research
      Citation Excerpt :

      Le Ber et al. (2009) simulated agricultural landscapes defined by two different tessellations (Voronoi and rectangular), i.e., the cover of the Euclidian plan by a countable number of geometric shapes, and two types of cropping pattern distributions (random or stochastic). Papaïx et al. (2014) developed a simple landscape generator that generates the polygon landscape mosaic based on a T-tessellation algorithm developed by Kiêu et al. (2013). Tessellation models have the advantage of being parametric, meaning that a set of parameters control the main features of the simulated landscapes.

    • Agricultural landscape generators for simulation models: A review of existing solutions and an outline of future directions

      2019, Ecological Modelling
      Citation Excerpt :

      A set of landscape pattern metrics is implemented to assess variations in landscape configuration and composition: patch cohesion, average patch shape, contagion, area weighted mean shape index, and landscape patch index. The simple landscape simulator within the Ddal framework generates a landscape mosaic based on the T-tessellation simulation algorithm developed by Kiêu et al. (2013) (Fig. 9). The input parameters number of fields, field surface average, field surface variability, and the square-like form of fields can be determined.

    • Spatial exposure-hazard and landscape models for assessing the impact of GM crops on non-target organisms

      2018, Science of the Total Environment
      Citation Excerpt :

      The landscape is first structured by partitioning a 5000 × 5000m2 domain with a Voronoi tessellation on I seed points drawn from a binomial homogeneous spatial point process(Chiu et al., 2013; Illian et al., 2008). More realistic tessellations could be used for drawing the geometry of agricultural fields (Kiêu et al., 2013), but it has been shown that Voronoi tessellations based on binomial point processes are relevant for assessing the interaction between the spatial arrangement of the landscape and the dispersal of pollen grains (Lavigne et al., 2008; Le Ber et al., 2009). Then, for a given proportion of GM fields p we draw both the spatial distribution of GM fields and host-margins on the set of polygons and segments defined by the tessellation using marked polygonal and point processes, and a thresholded stationary spatial Gaussian process with a Matérn covariance function whose range parameter ρ controls the spatial aggregation of emitting fields.

    • Perception-based foraging for competing resources: Assessing pest population dynamics at the landscape scale from heterogeneous resource distribution

      2015, Ecological Modelling
      Citation Excerpt :

      We consider this model as an accurate tool for studying the impacts of the spatial distribution of competitive resources, be they polygonal (as fields), linear (as hedgerows) or scattered across the landscape (in a less anthropic layout). On the other hand, the recent advances in agricultural landscape models (Kieu et al., 2013) enable to simulate a variety of field arrangements (and ensuing resource distributions for a given organism), as well as their sequential modifications towards realistic agricultural management policies. With such technical perspectives, light is expected to be shed on the mechanisms driving pest reactions to agricultural mosaics in forthcoming applications.

    • Pathogen population dynamics in agricultural landscapes: The Ddal modelling framework

      2014, Infection, Genetics and Evolution
      Citation Excerpt :

      These authors proposed a stochastic model that can be considered as a completely random distribution on the T-tessellation space and is closely related to a previous model introduced by Arak et al. (1993). Kiêu et al. (2013) improved the model by considering Gibbsian extensions based on an energy function so as to make it possible to control key LPIs of the field pattern (Fig. 2a–c). The simulation algorithm is based on the Metropolis–Hastings–Green principle that makes it possible to generate several field patterns with the same LPIs distribution (Fig. 2d–g).

    • Can epidemic control be achieved by altering landscape connectivity in agricultural systems?

      2014, Ecological Modelling
      Citation Excerpt :

      The field patterns are defined using a tessellation method that consists in filling the simulated landscape with polygons, controlling both the number of polygons and their geometrical shapes. The tessellation algorithm (Kiêu et al., 2013) is based on Markov Chain Monte-Carlo methods that make it possible to generate several field patterns with the same characteristics (Fig. 1b–d). Once the field pattern is established, the two varieties SV and RV are dispatched among the fields using a simulated annealing algorithm that performs constrained optimization and makes it possible to generate pseudo-random allocation replicates, while controlling the variety proportions and aggregation levels (Fig. 1b and d).

    View all citing articles on Scopus
    View full text