betaFeedback

WorksMathematics

Bayesian Wombling for Tracking Invasive Species

Abstract

This paper develops Bayesian wombling, or boundary analysis for tracking invasive species. We focus on three invasive species of concern in North America: Lycorma delicatula, Adelges tsugae, and Microstegium vimineum. We aim to demonstrate the applicability of wombling for supporting invasive pest management activities. In spatial science wombling is used widely to track regions of rapid change within the spatial reference domain. We use Gaussian process models for point-referenced spatial data followed by predictive inference on rates of change over curves using line integrals. We extend the wombling model based Bayesian inference framework to feature simple analytic forms that avoid high dimensional quadrature while offering uncertainty quantification over curvilinear spatial boundaries.

Introduction

Invasive alien species can have substantial biodiversity, environmental and economic impacts, with global impacts of invasive alien species estimated at more than $9 billion annually from 1970 - 2020 (see, e.g., Leroy et al., 2022). These impacts can be incredibly dependent on where the species invade, making species occurrence data and estimates of species distributions critical to being able to respond to, manage invasive species and devise interventions (see, e.g., Latzka et al., 2016). A common approach to managing invasive species applies surveillance based on where a species has been found, high risk areas of incursion, potential impacts, and species distribution models to guide efforts (see, e.g., Epanchin-Niell et al., 2010). In the case of species distribution models, these often estimate the potential area that a species could establish (i.e. potential niche; e.g. MAXENT Horrocks et al. [25], CLIMEX Taylor et al. [38], CLIMATCH Bomford [9], Bioclim Panetta and Dodd [33]), rather than its actual distribution (though see [30, 39] for examples of estimating actual distributions). Václavík and Meentemeyer [39] illustrate the importance of including constraints when trying to estimate actual species distributions. Therefore, estimating the boundaries of a species invasion would support effective management of invasive species under the typical conditions of resource constraints, by providing a more realistic determination of where to concentrate efforts.

At present, species distribution, phenology, and species distribution models are used to support risk assessment, risk management, and surveillance activities for invasive species (see, e.g., Srivastava et al., 2019; Hudgins et al., 2020; Belouard et al., 2025). These models generally provide a hypothesis of the potential range into which an invasive species can expand or create potential spread scenarios for planning management actions. While this information can guide surveillance and risk mitigation efforts, there is still considerable time spent determining optimal locations for pest survey sites to delimit and track the progress of an invasion, as well as determining whether a mitigation or management measure is effective at reducing the spread of a species. Identifying regions of rapid change in occurrence remains crucial for generating a hypothesis of the actual, rather than potential, distribution of an invasive species, as well as identifying spatial transitions that may be associated with invasion boundaries, barriers, or corridors.

Interpolated surfaces for logarithm of species occurrence counts for three invasive species

Figure 1. Interpolated surfaces for logarithm of species occurrence counts for (left) Adelges tsugae (center) Lycorma delicatula (right) Microstegium vimineum

Detecting regions of rapid change in spatial distribution of a species and subsequently the effect or presence of boundaries, commonly known as wombling (see, e.g. Womble, 1951; Gleyze et al., 2001; Banerjee, 2010), has been of ecological interest, for example, Holway [24] examines effects across a natural boundary in Argentine ants using standard analysis of variance, Fitzpatrick et al. [13] looks at boundaries in the spatial distribution of the hemlock woolly adelgid at the county-level, boundaries being shared regions for neighboring counties, for further examples (see, e.g. Gao et al., 2023; Wu and Banerjee, 2025). Wombling for point-referenced data, for instance, latitude-longitude or, easting-northing referenced data, demands substantially different methods, for example, Strydom and Poisot [37] considers formulating rates of change for woody plant distributions and looks for candidate boundaries only at sharp elevation changes, their spatial interpolation is deterministic. We are concerned with point-referenced Bayesian wombling under a hierarchical model (see, e.g., Banerjee et al., 2003; Banerjee and Gelfand, 2006; Qu et al., 2021; Halder et al., 2024) for the spatial distribution of an invasive species. Our choices of candidate boundaries are driven by high gradients in species distribution. Developing an easily-accessible framework that implements Bayesian wombling for use by the wider ecological community is faced with several challenges. Perhaps, the most severe being computational expense. Our statistical contributions lie in the use of analytic closed forms for posteriors that improves upon previous approaches while producing efficient Bayesian inference for hierarchical spatial models (see, e.g. Banerjee et al., 2014) and the novel application to tracking invasive species. We apply Bayesian wombling to detect boundaries in the distribution of three invasive species of concern in North America: Lycorma delicatula (LD) commonly known as the spotted lantern fly, Adelges tsugae (AT) commonly known as the hemlock woolly adelgid, and Microstegium vimineum (MV) commonly known as the Japanese stiltgrass, with the objective of demonstrating its applicability for supporting invasive pest and species management activities. The first two are insects and the third one is a plant species. There are no standard statistical approaches that offer model-based inference over boundaries (curves) within a spatial setting. However, when regions of interest are in question commonly used approaches are clustering (see, e.g., Ester et al., [11]) and hot-spot detection (see, e.g., Bivand and Wong, [8]). While wombling seeks predictive inference over curves offering fully model-based uncertainty quantification leveraging rates of change, clustering and hot-spot detection techniques aim to find homogeneous regions in the reference domain, unable to offer inference over curves separating such regions. We show detailed comparisons. We make note that the data considered here are spatial rather than spatiotemporal, these gradients characterize spatial changes in species occurrence and should not, by themselves, be interpreted as estimates of temporal spread. Rather, they identify spatial transitions that may be consistent with invasion fronts, barriers, corridors, or other features of the species distribution.

The manuscript evolves as follows: Section 2 elaborates on the processes that govern spatial rates of change followed by Section 3 which describes the underpinnings of Bayesian wombling and introduces the necessary results for computational improvements. Section 4 reviews Bayesian hierarchical models and posterior inference. Section 5 calibrates the performance of Bayesian wombling with closed forms on simulated data. Section 6 demonstrates the application to modeling invasive species counts, analyzing spatial rates of change in occurrence, and detecting boundaries. Section 7 concludes with a discussion and future avenues of research.

Spatial Derivative Processes

We consider {Y(s):s∈S⊆R2}\{Y(\mathbf{s}):\mathbf{s}\in\mathcal{S}\subseteq\mathbb{R}^{2}\} to be a univariate weakly stationary random field with zero mean and a positive definite covariance K(s,s′)=Cov⁡(Y(s),Y(s′))K(\mathbf{s},\mathbf{s}')=\operatorname{Cov}(Y(\mathbf{s}),Y(\mathbf{s}')) for locations s,s′∈R2\mathbf{s},\mathbf{s}'\in\mathbb{R}^{2}. Mean square smoothness (see, e.g., Stein, [36]) at an arbitrary location s0\mathbf{s}_{0} requires

Y(s0+hu)=Y(s0)+huT∂Y(s0)+h2u⊗2∂⊗2Y(s0)+o(h3∥u∥3),Y(\mathbf{s}_{0}+h\mathbf{u})=Y(\mathbf{s}_{0})+h\mathbf{u}^{\mathsf{T}}\partial Y(\mathbf{s}_{0})+h^{2}\mathbf{u}^{\otimes2}\partial^{\otimes2}Y(\mathbf{s}_{0})+o(h^{3}\lVert\mathbf{u}\rVert^{3}),

where, u=(u1,u2)T∈R2\mathbf{u}=(u_{1},u_{2})^{\mathsf{T}}\in\mathbb{R}^{2} is an arbitrary vector of directions, ∂\partial is the gradient operator, ∂Y(s0)=(∂∂xY(s0),∂∂yY(s0))T=(∂xY(s0),∂yY(s0))T\partial Y(\mathbf{s}_{0})=\left(\frac{\partial}{\partial x}Y(\mathbf{s}_{0}),\frac{\partial}{\partial y}Y(\mathbf{s}_{0})\right)^{\mathsf{T}}=(\partial_{x}Y(\mathbf{s}_{0}),\partial_{y}Y(\mathbf{s}_{0}))^{\mathsf{T}}, ⊗\otimes is the Kronecker vector product. Hence, u⊗2=(u12,u1u2,u2u1,u22)T\mathbf{u}^{\otimes2}=(u_{1}^{2},u_{1}u_{2},u_{2}u_{1},u_{2}^{2})^{\mathsf{T}} and ∂⊗2=∂⊗∂\partial^{\otimes2}=\partial\otimes\partial. Note that ∂⊗2Y(s0)\partial^{\otimes2}Y(\mathbf{s}_{0}) is the vectorized Hessian. The processes ∂Y(s0)\partial Y(\mathbf{s}_{0}) and ∂⊗2Y(s0)\partial^{\otimes2}Y(\mathbf{s}_{0}) govern rates of change in Y(s0)Y(\mathbf{s}_{0}). The gradient, or, first order rate of change, is captured by ∂Y(s0)\partial Y(\mathbf{s}_{0}) while, curvature is captured by ∂⊗2Y(s0)\partial^{\otimes2}Y(\mathbf{s}_{0}). Mean square differentiability (see Banerjee and Gelfand, [4]) of the first and second order for Y(s0)Y(\mathbf{s}_{0}) guarantees that uT∂Y(s0)\mathbf{u}^{\mathsf{T}}\partial Y(\mathbf{s}_{0}) and (u⊗v)T∂⊗2Y(s0)(\mathbf{u}\otimes\mathbf{v})^{\mathsf{T}}\partial^{\otimes2}Y(\mathbf{s}_{0}) are well-defined respectively, for any set of direction vectors u,v∈R2\mathbf{u},\mathbf{v}\in\mathbb{R}^{2}. We note that the entries of ∂⊗2Y(s0)\partial^{\otimes2}Y(\mathbf{s}_{0}) contain duplicates, both ∂xy2Y(s0)=∂2∂x∂yY(s0)\partial_{xy}^{2}Y(\mathbf{s}_{0})=\frac{\partial^{2}}{\partial x\partial y}Y(\mathbf{s}_{0}) and ∂yx2Y(s0)\partial_{yx}^{2}Y(\mathbf{s}_{0}) are included. To avoid singularities that arise from duplication, we work with ∂~⊗2Y(s0)=(∂xx2Y(s0),∂xy2Y(s0),∂yy2Y(s0))T\widetilde{\partial}^{\otimes2}Y(\mathbf{s}_{0})=\left(\partial_{xx}^{2}Y(\mathbf{s}_{0}),\partial_{xy}^{2}Y(\mathbf{s}_{0}),\partial_{yy}^{2}Y(\mathbf{s}_{0})\right)^{\mathsf{T}} comprised of only unique derivatives.

Statistical inference is devised for the joint process, L∗Y(s)=(Y(s),∂Y(s)T,∂⊗2Y(s)T)T\mathcal{L}^{*}Y(\mathbf{s})=\left(Y(\mathbf{s}),\partial Y(\mathbf{s})^{\mathsf{T}},\partial^{\otimes2}Y(\mathbf{s})^{\mathsf{T}}\right)^{\mathsf{T}}. Validity of the inference is considered at length in Banerjee et al. [6]; ); Halder et al. [19]. The process L∗Y(s)\mathcal{L}^{*}Y(\mathbf{s}) is also weakly stationary with a cross-covariance matrix,

∇L∗(Δ)=(K(Δ)∂K(Δ)T∂~2K(Δ)T−∂K(Δ)−∂2K(Δ)−∂~3K(Δ)T∂~2K(Δ)∂~3K(Δ)∂~4K(Δ)),(1)\nabla_{\mathcal{L}^{*}}(\Delta)=\begin{pmatrix} K(\Delta) & \partial K(\Delta)^{\mathsf{T}} & \widetilde{\partial}^{2}K(\Delta)^{\mathsf{T}}\\ -\partial K(\Delta) & -\partial^{2}K(\Delta) & -\widetilde{\partial}^{3}K(\Delta)^{\mathsf{T}}\\ \widetilde{\partial}^{2}K(\Delta) & \widetilde{\partial}^{3}K(\Delta) & \widetilde{\partial}^{4}K(\Delta) \end{pmatrix}, \tag*{(1)}

where Δ=s−s′\Delta=\mathbf{s}-\mathbf{s}', ∂K(Δ)\partial K(\Delta) is a 2×12\times1 vector of gradients, ∂~2K(Δ)\widetilde{\partial}^{2}K(\Delta) is a 3×13\times1 vector of unique curvatures, ∂~3K(Δ)\widetilde{\partial}^{3}K(\Delta) is a 3×23 \times2 matrix of third derivatives, ∂2K(Δ)\partial^{2}K(\Delta) is the 2×22 \times2 Hessian and ∂~4K(Δ)\widetilde{\partial}^{4}K(\Delta) is a 3×33 \times3 matrix of fourth order derivatives. Evidently, for V⋆(Δ)\mathbf{V}_{\star}(\Delta) in eq. (1) to be valid all entries need to be well-defined.

Let Y(s)∼GP(0,K(⋅;θ))\mathcal{Y}(s) \sim GP(0,K(\cdot;\theta)) denote a Gaussian process (GP) where K(Δ;θ)=Cov⁡(Y(s),Y(s′))K(\Delta;\theta) = \operatorname{Cov}(\mathcal{Y}(s),\mathcal{Y}(s')) with process parameters θ={σ2,ϕ}\theta= \{\sigma^{2},\phi\}. We will denote K(Δ;θ)=K(Δ)K(\Delta;\theta) = K(\Delta) to ease notation. The covariance function satisfies ∑i=1N∑j=1NaiajK(Δij)>0\sum_{i=1}^{N}\sum_{j=1}^{N}a_{i}a_{j}K(\Delta_{ij}) > 0 for any collection of coordinates {si:i=1,…,N}\{s_i : i = 1,\ldots,N\}. Under isotropy we have K(Δ)=K~(∥Δ∥)K(\Delta) = \widetilde{K}(\lVert\Delta\rVert). Let Y=(y(s1),…,y(sN))T\mathcal{Y} = (y(s_1),\ldots,y(s_N))^{\mathsf{T}} be the observed realization over S\mathcal{S}, Σy\Sigma_y be the N×NN \times N covariance matrix with entries K(si,sj)K(s_i,s_j), i,j=1,…,Ni,j = 1,\ldots,N and s0s_0 an arbitrary location of interest. The joint distribution is as follows:

Y,∂Y(s0),∂~2Y(s0)∣θ∼NN+5(0N+5,(ΣyK1K2−K1T−∂2K(0)−∂~3K(0)TK2T∂~3K(0)∂~4K(0))),(2)\mathcal{Y},\partial\mathcal{Y}(s_0),\widetilde{\partial}^{2}\mathcal{Y}(s_0)\mid\theta\sim\mathcal{N}_{N+5}\left(0_{N+5},\left(\begin{array}{ccc} \Sigma_y & \mathbf{K}_1 & \mathbf{K}_2 \\ -\mathbf{K}_1^{\mathsf{T}} & -\partial^{2}K(0) & -\widetilde{\partial}^{3}K(0)^{\mathsf{T}} \\ \mathbf{K}_2^{\mathsf{T}} & \widetilde{\partial}^{3}K(0) & \widetilde{\partial}^{4}K(0) \end{array}\right)\right), \tag*{(2)}

where, Nd\mathcal{N}_d denotes the dd-variate Gaussian distribution, K1=(∂K(δ10)T,…,∂K(δN0)T)T\mathbf{K}_1 = (\partial K(\delta_{10})^{\mathsf{T},\ldots,\partial K(\delta_{N0})^{\mathsf{T}})^{\mathsf{T}}}, K2=(∂~2K(δ10)T,…,∂~2K(δN0)T)T\mathbf{K}_2 = (\widetilde{\partial}^{2}K(\delta_{10})^{\mathsf{T}},\ldots,\widetilde{\partial}^{2}K(\delta_{N0})^{\mathsf{T}})^{\mathsf{T}} and δi0=si−s0\delta_{i0} = s_i-s_0, i=1,…,Ni = 1,\ldots,N. The posterior predictive distribution for rates of change at s0s_0 is P(∂Y(s0),∂~2Y(s0)∣Y)=∫P(∂Y(s0),∂~2Y(s0)∣Y,θ) P(θ∣Y) dθP(\partial\mathcal{Y}(s_0),\widetilde{\partial}^{2}\mathcal{Y}(s_0)\mid\mathcal{Y}) = \int P(\partial\mathcal{Y}(s_0),\widetilde{\partial}^{2}\mathcal{Y}(s_0)\mid\mathcal{Y},\theta)\,P(\theta\mid\mathcal{Y})\,d\theta. Posterior sampling proceeds in a one-for-one fashion corresponding to posterior samples of θ\theta. From eq. (2) the resulting full conditional distribution is obtained as follows:

\begin{pmatrix} \partial\mathcal{Y}(s_0)\\ \widetilde{\partial}^{2}\mathcal{Y}(s_0) \end{pmatrix} \mid\mathcal{Y} \sim\mathcal{N}_{5}\left(-\begin{pmatrix}\mathbf{K}_1\\\mathbf{K}_2\end{pmatrix}^{\mathsf{T}}\Sigma_y^{-1}\mathcal{Y}, \begin{pmatrix} -\partial^{2}K(0) & -\widetilde{\partial}^{3}K(0)^{\mathsf{T}}\\ \widetilde{\partial}^{3}K(0) & \widetilde{\partial}^{4}K(0) \end{pmatrix} -\begin{pmatrix}\mathbf{K}_1\\\mathbf{K}_2\end{pmatrix}^{\mathsf{T}}\Sigma_y^{-1}\begin{pmatrix}\mathbf{K}_1\\\mathbf{K}_2\end{pmatrix}\right). \tag*{(3)}\end{pmatrix}

We use the Matérn class of kernels, K(∥Δ∥,θ)=σ2Γ(ν)−121−ν(2νϕ∥Δ∥)νKν(2νϕ∥Δ∥)K(\lVert\Delta\rVert,\theta) = \sigma^{2}\Gamma(\nu)^{-1}2^{1-\nu}\left(\sqrt{2\nu}\phi\lVert\Delta\rVert\right)^{\nu}K_{\nu}\left(\sqrt{2\nu}\phi\lVert\Delta\rVert\right), where Kν(⋅)K_{\nu}(\cdot) is the modified Bessel function of the second kind [1] featuring a fractal parameter ν\nu that controls process smoothness, a spatial range parameter ϕ\phi and an overall variance parameter σ2\sigma^{2}. Note that the GP needs to be sufficiently smooth to ensure the existence of all differential processes, specifically, ν>2\nu> 2 ensures the existence of derivative and the curvature process.

Bayesian Wombling

In general wombling seeks posterior predictive inference on line integrals over curves

Γ(C)=(∫CuT∂Y(s) ds,∫Cu⊗2∂~2Y(s) ds)T,(4)\mathbf{\Gamma}(\mathcal{C}) = \left(\int_{\mathcal{C}} \mathbf{u}^{\mathsf{T}}\partial\mathcal{Y}(s)\,ds,\int_{\mathcal{C}}\mathbf{u}^{\otimes2}\widetilde{\partial}^{2}\mathcal{Y}(s)\,ds\right)^{\mathsf{T}}, \tag*{(4)}

where C\mathcal{C} is a curve of interest to the investigator (see, e.g., [19, 5 Sec. 3]. A curve, C\mathcal{C} is said to form a wombling boundary if the total, or average wombling measures, which are defined as Γ‾(C)=Γ(C)/ℓ(C)\overline{\mathbf{\Gamma}}(\mathcal{C}) = \mathbf{\Gamma}(\mathcal{C})/\ell(\mathcal{C}), where ℓ\ell is the arc-length measure, are significantly large. The curves can be topologically open or closed, we use ∫\int or ∮\oint accordingly in eq. (4). Selecting an appropriate direction, u\mathbf{u} becomes crucial when measuring rates of change. The curve C\mathcal{C} typically tracks a region of rapid change in the spatial domain and hence, the direction normal to C\mathcal{C} is of importance. We denote the normal to C\mathcal{C} at ss by n(s)\mathbf{n}(s) and set u=n(s)\mathbf{u} = \mathbf{n}(s) in the line integrals of eq. (4). The curve C\mathcal{C} is deemed to be a wombling boundary with regard to the gradient (and/or curvature) if the first (and/or second) element of Γ(C)\mathbf{\Gamma}(\mathcal{C}) is large. Focusing on the choices for C\mathcal{C}, clearly not all curves ensure the existence of n(s)\mathbf{n}(s) at every ss. Parametric smooth curves offer some respite in that regard. We work with C={s(t)=(s1(t),s2(t)):t∈I⊂R}\mathcal{C} = \{s(t) = (s_1(t),s_2(t)) : t \in\mathcal{I} \subset\mathbb{R}\}. As tt varies over I\mathcal{I}, s(t)s(t) traces out C\mathcal{C}. We assume ∥s′(t)∥≠0\lVert s'(t)\rVert\ne0 which ensures n(s)=∥s′(t)∥−1(s2′(t),−s1′(t))Tn(s) = \lVert s'(t)\rVert^{-1}(s_2'(t),-s_1'(t))^{\mathsf{T}} is well-defined.

The arc-length, ℓ(C)=∫I∥s′(t)∥ dt\ell(\mathcal{C}) = \int_{\mathcal{I}} \lVert s'(t)\rVert\,dt. For parametric curves Γ(C)\Gamma(\mathcal{C}) can be expressed as,

Γ(C)=(∫In(s(t))T∂Y(s(t))∥s′(t)∥ dt,∫In(s(t))⊗2∂⊗2Y(s(t))∥s′(t)∥ dt)T.\Gamma(\mathcal{C}) = \left(\int_{\mathcal{I}} n(s(t))^{\mathsf{T}}\partial Y(s(t))\lVert s'(t)\rVert\,dt,\int_{\mathcal{I}} n(s(t))^{\otimes2}\partial^{\otimes2}Y(s(t))\lVert s'(t)\rVert\,dt\right)^{\mathsf{T}}.

Let I=[0,t∗]\mathcal{I} = [0,t^*], and the curve traced out over I\mathcal{I} be denoted as Ct∗\mathcal{C}_{t^*}. Statistical inference for Γ(Ct∗)\Gamma(\mathcal{C}_{t^*}) follows from ∂Y(s)\partial Y(s) and ∂2Y(s)\partial^2Y(s) being GPs, as seen in eq. (3), Γ(Ct∗)∼N2(02,KΓ(t∗,t∗))\Gamma(\mathcal{C}_{t^*}) \sim\mathcal{N}_2(0_2,\mathbf{K}_{\Gamma}(t^*,t^*)), where KΓ(t∗,t∗)\mathbf{K}_{\Gamma}(t^*,t^*) is a 2×22 \times2 matrix with entries

kij(t∗,t∗)=(−1)i∫0t∗∫0t∗ai(t1)T∂i+jK(Δ(t1,t2))aj(t2)∥s′(t1)∥∥s′(t2)∥ dt1 dt2,(5)k_{ij}(t^*,t^*) = (-1)^i \int_0^{t^*}\int_0^{t^*} a_i(t_1)^{\mathsf{T}}\partial^{i+j}K(\Delta(t_1,t_2))a_j(t_2)\lVert s'(t_1)\rVert\lVert s'(t_2)\rVert\,dt_1\,dt_2, \tag*{(5)}

where a1(t)=n(s(t))a_1(t) = n(s(t)), a2(t)=E2n(s(t))⊗2a_2(t) = \mathcal{E}_2 n(s(t))^{\otimes2}, with E2=(100010001)\mathcal{E}_2 = \begin{pmatrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{pmatrix} being an elimination matrix and Δ(t1,t2)=s(t2)−s(t1)\Delta(t_1,t_2) = s(t_2) - s(t_1) for i,j=1,2i,j = 1,2. Predictive inference on Γ(C)\Gamma(\mathcal{C}) requires

Y,Γ(Ct∗)∣θ∼NN+2(0N+2,[ΣYGΓ(t∗)TGΓ(t∗)KΓ(t∗,t∗)]),(6)\mathbf{Y},\Gamma(\mathcal{C}_{t^*}) \mid\theta\sim\mathcal{N}_{N+2}\left(0_{N+2},\begin{bmatrix} \boldsymbol{\Sigma}_{Y} & \mathcal{G}_{\Gamma}(t^*)^{\mathsf{T}} \\ \mathcal{G}_{\Gamma}(t^*) & \mathbf{K}_{\Gamma}(t^*,t^*) \end{bmatrix}\right), \tag*{(6)}

where GΓ(t∗)=(γ1(t∗)T⋮γN(t∗)T)\mathcal{G}_{\Gamma}(t^*) = \begin{pmatrix} \gamma_1(t^*)^{\mathsf{T}} \\ \vdots\\ \gamma_N(t^*)^{\mathsf{T}} \end{pmatrix} is an N×2N \times2 matrix with entries

γi(t∗)=(∫0t∗n(s(t))T∂K(Δj(t))∥s′(t)∥ dt,∫0t∗n(s(t))⊗2∂⊗2K(Δj(t))∥s′(t)∥ dt)T,(7)\gamma_i(t^*) = \left(\int_0^{t^*} n(s(t))^{\mathsf{T}}\partial K(\Delta_j(t))\lVert s'(t)\rVert\,dt,\int_0^{t^*} n(s(t))^{\otimes2}\partial^{\otimes2}K(\Delta_j(t))\lVert s'(t)\rVert\,dt\right)^{\mathsf{T}}, \tag*{(7)}

where Δj(t)=s(t)−sj\Delta_j(t) = s(t) - s_j. Posterior predictive inference proceeds one-for-one (similar to eq. (3)) using Γ(Ct∗)∣Y∼N2(−GΓ(t∗)ΣY−1Y,KΓ(t∗,t∗)−GΓ(t∗)ΣY−1GΓ(t∗)T)\Gamma(\mathcal{C}_{t^*}) \mid\mathbf{Y} \sim\mathcal{N}_2(-\mathcal{G}_{\Gamma}(t^*)\boldsymbol{\Sigma}_Y^{-1}\mathbf{Y},\mathbf{K}_{\Gamma}(t^*,t^*)-\mathcal{G}_{\Gamma}(t^*)\boldsymbol{\Sigma}_Y^{-1}\mathcal{G}_{\Gamma}(t^*)^{\mathsf{T}}).

In practice, modern computing environments store curves as a set of points. As a result, it suffices to demonstrate wombling for rectilinear approximations to smooth curves where predictive inference is performed iteratively on segments. We show the inference for one generic segment. Let C={s(t0),s(t1),…,s(tnp)}\mathcal{C} = \{s(t_0),s(t_1),\ldots,s(t_{n_p})\}, then the ii-th segment, Cti={s(t)=s(ti−1)+tui:t∈[0,ti]}\mathcal{C}_{t_i} = \{s(t) = s(t_{i-1}) + t u_i : t \in[0,t_i]\}, ti=∥s(ti)−s(ti−1)∥t_i = \lVert s(t_i)-s(t_{i-1})\rVert and ui=ti−1(s(ti)−s(ti−1))u_i = t_i^{-1}(s(t_i)-s(t_{i-1})). Clearly, ∥ui∥=1\lVert u_i\rVert= 1, ∥s′(t)∥=1\lVert s'(t)\rVert= 1 and the normal to Cti\mathcal{C}_{t_i} is ui⊥=(ui2,−ui1)Tu_i^\perp= (u_{i2},-u_{i1})^{\mathsf{T}}. For predictive inference on Γ(Cti)\Gamma(\mathcal{C}_{t_i}) we set Δ(t1,t2)=(t2−t1)ui\Delta(t_1,t_2) = (t_2-t_1)u_i in eq. (5) and Δj(t)=Δi−1,j+tui=(si−1−sj)+tui−1\Delta_j(t) = \Delta_{i-1,j} + t u_i = (s_{i-1}-s_j) + t u_{i-1} in eq. (7).

Closed Forms

We acknowledge that eq. (5) requires 2-dimensional quadrature which is computationally expensive to evaluate. Conventional approaches to wombling (see, e.g. [5, 19]) evaluate eq. (5) using two-dimensional quadrature using a Matérn kernel. We improve on this by showing that closed form analytic expressions exist for the entries of KΓ(t∗,t∗)\mathbf{K}_{\Gamma}(t^*,t^*). The benefits are reduced computation time and ease of implementation for which specific details are in Section 5. The following results provide more details.

Theorem 1. On each rectilinear segment, Ct∗={s0+tu:t∈[0,t∗]}\mathcal{C}_{t^*} = \{s_0 + t u : t \in[0,t^*]\}, where uu is a unit vector and u⊥u^\perp is its normal, considering the terms in the cross-covariance matrix we have

kij(t∗,t∗)=(−1)i∫0t∗∫0t∗ai(t1)T∂i+jK(Δ(t1,t2))aj(t2)∥s′(t1)∥∥s′(t2)∥ dt1 dt2,k_{ij}(t^*,t^*) = (-1)^i \int_0^{t^*}\int_0^{t^*} a_i(t_1)^{\mathsf{T}}\partial^{i+j}K(\Delta(t_1,t_2))a_j(t_2)\lVert s'(t_1)\rVert\lVert s'(t_2)\rVert\,dt_1\,dt_2,
=(−1)i∫−t∗t∗(t∗−∣x∣)aiT∂i+jK(xu)aj dx,i,j=1,2,= (-1)^i \int_{-t^*}^{t^*}(t^*-\lvert x\rvert)a_i^{\mathsf{T}}\partial^{i+j}K(xu)a_j\,dx,\quad i,j=1,2,

where a1(t)=n(s(t))a_1(t)=n(s(t)) is the normal to the segment, a2(t)=E2n(s(t))⊗n(s(t))a_2(t)=\mathcal{E}_2 n(s(t))\otimes n(s(t)), where

E2=(11111)\mathcal{E}_2=\begin{pmatrix}1&1&1\\&1&\\&&1\end{pmatrix}

is an elimination matrix and Δ(t1,t2)=s(t2)−s(t1)\Delta(t_1,t_2)=s(t_2)-s(t_1).

Proof. Considering the parametric segment, Ct∗={s0+tu:t∈[0,t∗]}C_{t^*}=\{s_0+tu:t\in[0,t^*]\}, where u=(u1,u2)Tu=(u_1,u_2)^{\mathsf T} is a unit vector, ∥u∥=1\lVert u\rVert=1, u⊥u^\perp is its normal, uTu⊥=0u^{\mathsf T}u^\perp=0, a1(t)=a1=u⊥a_1(t)=a_1=u^\perp and a2(t)=a2=E2(u⊥⊗u⊥)a_2(t)=a_2=\mathcal{E}_2(u^\perp\otimes u^\perp) are free of tt, and Δ(t1,t2)=(t2−t1)u\Delta(t_1,t_2)=(t_2-t_1)u. The integrand in kij(t∗,t∗)=(−1)i∫0t∗∫0t∗aiT∂i+jK((t2−t1)u)aj dt1 dt2k_{ij}(t^*,t^*)=(-1)^i\int_0^{t^*}\int_0^{t^*}a_i^{\mathsf T}\partial^{i+j}K((t_2-t_1)u)a_j\,dt_1\,dt_2, depends only on (t2−t1)(t_2-t_1). Making a change of variable—define x=t2−t1x=t_2-t_1, y=t2+t1y=t_2+t_1, implying x+y2=t2\frac{x+y}{2}=t_2 and y−x2=t1\frac{y-x}{2}=t_1. The Jacobian is 12\frac{1}{2}. Hence, 0≤y+x≤2t∗0\le y+x\le2t^* and 0≤y−x≤2t∗0\le y-x\le2t^*. This implies ∣x∣≤y≤2t∗−∣x∣\lvert x\rvert\le y\le2t^*-\lvert x\rvert and −t∗≤x≤t∗-t^*\le x\le t^*. Making the substitution above reduces,

kij(t∗,t∗)=(−1)i2∫−t∗t∗∫∣x∣2t∗−∣x∣aiT∂i+jK(xu)aj dy dx=(−1)i∫−t∗t∗(t∗−∣x∣)aiT∂i+jK(xu)aj dx.k_{ij}(t^*,t^*)=\frac{(-1)^i}{2}\int_{-t^*}^{t^*}\int_{\lvert x\rvert}^{2t^*-\lvert x\rvert}a_i^{\mathsf T}\partial^{i+j}K(xu)a_j\,dy\,dx=(-1)^i\int_{-t^*}^{t^*}(t^*-\lvert x\rvert)a_i^{\mathsf T}\partial^{i+j}K(xu)a_j\,dx.

Setting i=j=1i=j=1, produces the scenario in Banerjee and Gelfand [5].

The Matérn covariance has analytic expressions for specific values of the fractal parameter,

K(∥Δ∥,θ)={σ2(1+3ϕ∥Δ∥)e−3ϕ∥Δ∥,ν=3/2,σ2(1+5ϕ∥Δ∥+53ϕ2∥Δ∥2)e−5ϕ∥Δ∥,ν=5/2.(8)K(\lVert\Delta\rVert,\theta)= \begin{cases} \sigma^2(1+\sqrt{3}\phi\lVert\Delta\rVert)e^{-\sqrt{3}\phi\lVert\Delta\rVert},&\nu=3/2,\\ \sigma^2\left(1+\sqrt{5}\phi\lVert\Delta\rVert+\frac{5}{3}\phi^2\lVert\Delta\rVert^2\right)e^{-\sqrt{5}\phi\lVert\Delta\rVert},&\nu=5/2. \end{cases} \tag*{(8)}

For ν=3/2\nu=3/2 it is once mean square differentiable and for ν=5/2\nu=5/2 it is twice mean square differentiable at 0.

Theorem 2. For the Matérn kernel, with ν=3/2\nu=3/2 in eq. (8), the variance of the gradient based wombling measure is k11(t∗,t∗)=2σ2(3ϕt∗−1+e−3ϕt∗)k_{11}(t^*,t^*)=2\sigma^2\left(\sqrt{3}\phi t^*-1+e^{-\sqrt{3}\phi t^*}\right). In case ν=5/2\nu=5/2 in eq. (8), the cross-covariance matrix for the wombling measures based on spatial gradient and curvature requires k11(t∗,t∗)=2σ23(25ϕt∗−3+(5ϕt∗+3)e−5ϕt∗)k_{11}(t^*,t^*)=\frac{2\sigma^2}{3}\left(2\sqrt{5}\phi t^*-3+(\sqrt{5}\phi t^*+3)e^{-\sqrt{5}\phi t^*}\right), k22(t∗,t∗)=10σ2ϕ2(5ϕt∗−1+e−5ϕt∗)k_{22}(t^*,t^*)=10\sigma^2\phi^2\left(\sqrt{5}\phi t^*-1+e^{-\sqrt{5}\phi t^*}\right) and k21(t∗,t∗)=−k12(t∗,t∗)=0k_{21}(t^*,t^*)=-k_{12}(t^*,t^*)=0. For the Gaussian kernel,

VΓ(t∗)=2σ2[πϕt∗{2Φ(2ϕt∗)−1}−(1−e−ϕ2t∗2)](1006ϕ2),V_{\Gamma}(t^*)=2\sigma^2\left[\sqrt{\pi}\phi t^*\left\{2\Phi\left(\sqrt{2}\phi t^*\right)-1\right\}-\left(1-e^{-\phi^2t^{*2}}\right)\right]\begin{pmatrix}1&0\\0&6\phi^2\end{pmatrix},

where Φ(⋅)\Phi(\cdot) is the cdf for the standard Gaussian probability density.

Proof. Using Theorem 1, if ν=3/2\nu=3/2, then we obtain k11(t∗,t∗)=3σ2ϕ2∫−t∗t∗(t∗−∣x∣)e−3ϕ∣x∣ dx=2σ2(3ϕt∗−1+e−3ϕt∗)k_{11}(t^*,t^*)=3\sigma^2\phi^2\int_{-t^*}^{t^*}(t^*-\lvert x\rvert)e^{-\sqrt{3}\phi\lvert x\rvert}\,dx=2\sigma^2\left(\sqrt{3}\phi t^*-1+e^{-\sqrt{3}\phi t^*}\right). If ν=5/2\nu=5/2, then k11(t∗,t∗)=53σ2ϕ2∫−t∗t∗(t∗−∣x∣)(1+5ϕ∣x∣)e−5ϕ∣x∣ dx=2σ23(25ϕt∗−3+(5ϕt∗+3)e−5ϕt∗)k_{11}(t^*,t^*)=\frac{5}{3}\sigma^2\phi^2\int_{-t^*}^{t^*}(t^*-\lvert x\rvert)(1+\sqrt{5}\phi\lvert x\rvert)e^{-\sqrt{5}\phi\lvert x\rvert}\,dx=\frac{2\sigma^2}{3}\left(2\sqrt{5}\phi t^*-3+(\sqrt{5}\phi t^*+3)e^{-\sqrt{5}\phi t^*}\right). For terms k21k_{21} and k12k_{12}, let us consider a1T∂3K(xu)a2=2553σ2ϕ3e−5ϕ∣x∣∣x∣{a1TA1a2−5ϕ∣x∣a1TA2a2}a_1^{\mathsf T}\partial^3K(xu)a_2=\frac{25\sqrt{5}}{3}\sigma^2\phi^3e^{-\sqrt{5}\phi\lvert x\rvert}\lvert x\rvert\{a_1^{\mathsf T}A_1a_2-\sqrt{5}\phi\lvert x\rvert a_1^{\mathsf T}A_2a_2\}, where

A1=(3u1u2u1u2u13u2),A2=(u13u12u2u22u1u12u2u22u1u23).A_1=\begin{pmatrix}3u_1u_2&u_1\\u_2&u_1&3u_2\end{pmatrix},\qquad A_2=\begin{pmatrix}u_1^3&u_1^2u_2&u_2^2u_1\\u_1^2u_2&u_2^2u_1&u_2^3\end{pmatrix}.

The first and the second term both reduce to 0. Hence, k21(t∗,t∗)=−k12(t∗,t∗)=0k_{21}(t^*,t^*)=-k_{12}(t^*,t^*)=0. Next,

a2T∂4K(xu)a2=253σ2ϕ4e−5ϕ∣x∣{a2TA3a2−5ϕ∣x∣a2TA4(5ϕ∣x∣+1)a2},a_2^{\mathsf T}\partial^4K(xu)a_2=\frac{25}{3}\sigma^2\phi^4e^{-\sqrt{5}\phi\lvert x\rvert}\{a_2^{\mathsf T}A_3a_2-\sqrt{5}\phi\lvert x\rvert a_2^{\mathsf T}A_4(\sqrt{5}\phi\lvert x\rvert+1)a_2\},

where

A3=(301010103),A4(x′)=(6u22−x′u14(3−x′u22)u1u21−x′u22u22(3−x′u22)u1u21−x′u12u22(3−x′u22)u1u21−x′u12u22(3−x′u22)u1u26u22−x′u14).A_3=\begin{pmatrix}3&0&1\\0&1&0\\1&0&3\end{pmatrix},\qquad A_4(x')=\begin{pmatrix} 6u_2^2-x'u_1^4&(3-x'u_2^2)u_1u_2&1-x'u_2^2u_2^2\\ (3-x'u_2^2)u_1u_2&1-x'u_1^2u_2^2&(3-x'u_2^2)u_1u_2\\ 1-x'u_1^2u_2^2&(3-x'u_2^2)u_1u_2&6u_2^2-x'u_1^4 \end{pmatrix}.

After some algebra, a2T∂4K(xu)a2=25σ2ϕ4e−5ϕ∣x∣{1−25ϕ∣x∣u1⊥u2⊥(u1u2+u1⊥u2⊥)}a_2^{\mathsf T}\partial^4K(xu)a_2=25\sigma^2\phi^4e^{-\sqrt{5}\phi\lvert x\rvert}\{1-2\sqrt{5}\phi\lvert x\rvert u_1^\perp u_2^\perp(u_1u_2+u_1^\perp u_2^\perp)\}. Observe that u1⊥=u2u_1^\perp=u_2 and u2⊥=−u1u_2^\perp=-u_1, substituting we get k22(t∗,t∗)=10σ2ϕ2(5ϕt∗−1+e−5ϕt∗)k_{22}(t^*,t^*)=10\sigma^2\phi^2\left(\sqrt{5}\phi t^*-1+e^{-\sqrt{5}\phi t^*}\right).

For the Gaussian kernel following the same steps we obtain, k11(t∗,t∗)=2σ2ϕ2∫−t∗t∗(t∗−∣x∣)e−ϕ2x2 dx=2σ2[πϕt∗{2Φ(2ϕt∗)−1}−(1−e−ϕ2t∗2)]k_{11}(t^{*},t^{*}) = 2\sigma^{2}\phi^{2}\int_{-t^{*}}^{t^{*}}(t^{*}-|x|)e^{-\phi^{2}x^{2}}\,dx = 2\sigma^{2}\left[\sqrt{\pi}\phi t^{*}\left\{2\Phi\left(\sqrt{2}\phi t^{*}\right)-1\right\}-\left(1-e^{-\phi^{2}t^{*2}}\right)\right]. Note that, similar to before, a1T∂3K(xu) a2=4σ2ϕ4e−ϕ2x2{a1TA1a2−2ϕ2x2 a1TA2a2}a_{1}^{\mathsf{T}}\partial^{3}K(xu)\,a_{2}=4\sigma^{2}\phi^{4}e^{-\phi^{2}x^{2}}\left\{a_{1}^{\mathsf{T}}A_{1}a_{2}-2\phi^{2}x^{2}\,a_{1}^{\mathsf{T}}A_{2}a_{2}\right\}, the first and second terms equate to 0 implying k21(t∗,t∗)=−k12(t∗,t∗)=0k_{21}(t^{*},t^{*})=-k_{12}(t^{*},t^{*})=0. Next, a2T∂4K(xu) a2=4σ2ϕ4e−ϕ2x2{a2TA3 a2−2ϕ2x2a2TA4(2ϕ2x2) a2}a_{2}^{\mathsf{T}}\partial^{4}K(xu)\,a_{2}=4\sigma^{2}\phi^{4}e^{-\phi^{2}x^{2}}\left\{a_{2}^{\mathsf{T}}A_{3}\,a_{2}-2\phi^{2}x^{2}a_{2}^{\mathsf{T}}A_{4}(2\phi^{2}x^{2})\,a_{2}\right\} yielding, a2T∂4K(xu) a2=12σ2ϕ4e−ϕ2x2{1−4ϕ2x2u1⊥u2⊥(u1u2+u1⊥u2⊥)}a_{2}^{\mathsf{T}}\partial^{4}K(xu)\,a_{2}=12\sigma^{2}\phi^{4}e^{-\phi^{2}x^{2}}\left\{1-4\phi^{2}x^{2}u_{1}^{\perp}u_{2}^{\perp}(u_{1}u_{2}+u_{1}^{\perp}u_{2}^{\perp})\right\}. Substituting, we get k22(t∗,t∗)=6ϕ2k11k_{22}(t^{*},t^{*})=6\phi^{2}k_{11}. □

Theorem 2 interestingly shows that the posited cross-covariance matrix, VΓ(t∗)\mathbf{V}_{\Gamma}(t^{*}), reduces to a variance-covariance matrix. Note that Theorems 1 and 2 are valid only for ν=3/2\nu=3/2, ν=5/2\nu=5/2 and ν→∞\nu\to\infty (the squared exponential) kernel. Compared with a Matérn kernel featuring ν=52\nu=\frac{5}{2}, the matrix has a simpler form for the Gaussian case. Finally, considering the covariance between Z(si)Z(s_i), i=1,…,Ni=1,\ldots,N and the wombling measures—in the Gaussian case, closed-forms are available (see, e.g., Halder et al., [19], end of Section 3). The inferential exercise of wombling does not require quadrature when using a Gaussian kernel however, only one-dimensional quadrature is required for the same when using a Matérn kernel.

Bayesian Hierarchical Models

A Bayesian hierarchical model is specified as follows:

Y(s)=μ(s,β)+Z(s)+ϵ(s),(9)Y(s)=\mu(s,\beta)+Z(s)+\epsilon(s), \tag*{(9)}

where μ(s,β)=x(s)Tβ\mu(s,\beta)=x(s)^{\mathsf{T}}\beta, Z(s)∼GP(0,K(⋅;σ2,ϕ))Z(s)\sim GP(0,K(\cdot;\sigma^{2},\phi)) and ϵ(s)∼N(0,τ2)\epsilon(s)\sim N(0,\tau^{2}) is a white noise process. The process parameters are θ={σ2,ϕ,τ2}\theta=\{\sigma^{2},\phi,\tau^{2}\}. A customary collapsed posterior (see, e.g. Finley et al., [12]) for θ\theta is specified as follows:

P(θ∣Y)∝U(ϕ∣aϕ,bϕ)×IG(σ2∣aσ,bσ)×IG(τ2∣aτ,bτ)×NN(Y∣Xβ,Σ+τ2IN),(10)P(\theta\mid\mathbf{Y})\propto U(\phi\mid a_{\phi},b_{\phi})\times IG(\sigma^{2}\mid a_{\sigma},b_{\sigma})\times IG(\tau^{2}\mid a_{\tau},b_{\tau})\times\mathcal{N}_{N}(\mathbf{Y}\mid X\beta,\boldsymbol{\Sigma}+\tau^{2}I_{N}), \tag*{(10)}

where Σ=σ2RZ(ϕ)\boldsymbol{\Sigma}=\sigma^{2}\mathbf{R}_{Z}(\phi), with RZ(ϕ)\mathbf{R}_{Z}(\phi) being the correlation matrix corresponding to K(⋅;σ2,ϕ)K(\cdot;\sigma^{2},\phi), U(⋅∣⋅)U(\cdot\mid\cdot) is the uniform distribution and IG(⋅∣⋅)IG(\cdot\mid\cdot) is the inverse-gamma distribution. Hyperparameters are chosen such that the resulting prior on θ\theta has a large variance, being “non-informative” in that sense. For example, common choices for aϕ=10−3a_{\phi}=10^{-3} and bϕ=300b_{\phi}=300, where aϕa_{\phi} can be reduced and bϕb_{\phi} can be increased if the posterior estimate of ϕ\phi from initial pilot runs is close to either one; aσ=bσ=aτ=bτ=2a_{\sigma}=b_{\sigma}=a_{\tau}=b_{\tau}=2 so as to generate an IGIG distribution with mean of 2 and infinite variance. Posterior samples for Z=(Z(s1),…,Z(sn))T\mathbf{Z}=(Z(s_{1}),\ldots,Z(s_{n}))^{\mathsf{T}} and β\beta are generated one-for-one corresponding to posterior samples of θ\theta using a Gibbs sampling scheme. For more details on the conditional distribution, P(θ∣Y)P(\theta\mid\mathbf{Y}) the interested reader is referred to Halder et al. [19], § 4.

Note that, our discussion on differential processes has been solely on the process, Z(s)Z(s), these processes can be associated with Y(s)Y(s), provided μ(s,β)\mu(s,\beta) admits spatial smoothness and, the location at which we are interested in obtaining inference on gradients is not one at which the process is observed. Predictive inference is sought for LTZ(s0)\mathcal{L}^{\mathsf{T}}Z(s_{0}) at an arbitrary location s0s_{0} through

P(∂Z(s0)T,∂~2Z(s0)T∣Y)=∫P(∂Z(s0)T,∂~2Z(s0)T∣Z,θ) P(Z∣Y,θ) P(θ∣Y) dθ dZ.(11)P(\partial Z(s_{0})^{\mathsf{T}},\widetilde{\partial}^{2}Z(s_{0})^{\mathsf{T}}\mid\mathbf{Y})=\int P(\partial Z(s_{0})^{\mathsf{T}},\widetilde{\partial}^{2}Z(s_{0})^{\mathsf{T}}\mid\mathcal{Z},\theta)\,P(\mathcal{Z}\mid\mathbf{Y},\theta)\,P(\theta\mid\mathbf{Y})\,d\theta\,d\mathcal{Z}. \tag*{(11)}

Similarly, for the wombling measures, inference proceeds using

P(ΓZ(Ct∗)∣Y)=∫P(ΓZ(Ct∗)∣Z,θ) P(Z∣Y,θ) P(θ∣Y) dθ dZ.(12)P(\Gamma_{Z}(C_{t^{*}})\mid\mathbf{Y})=\int P(\Gamma_{Z}(C_{t^{*}})\mid\mathcal{Z},\theta)\,P(\mathcal{Z}\mid\mathbf{Y},\theta)\,P(\theta\mid\mathbf{Y})\,d\theta\,d\mathcal{Z}. \tag*{(12)}

Simulation Experiments

Gradients and curvatures are not observed in reality rather they are induced by a parent spatial process. For the purpose of calibrating our inference, we use data that is not generated by the model in eq. (9) in our simulation experiments. We produce data using patterns that yield closed form expressions for rates of change. We generate N=100N = 100 observations over [−10,10]×[−10,10]⊆R2[-10, 10] \times[-10, 10] \subseteq\mathbb{R}^{2} arising from, y(s)∼N1(μ0(s)=20sin⁡∥s∥,τ2)y(s) \sim\mathcal{N}_{1}(\mu_{0}(s) = 20 \sin\lVert s \rVert, \tau^{2}), setting τ2=1\tau^{2} = 1. The patterned data is shown in the left plot, first row of Figure 2.

Interpolated surfaces for data, true and estimated gradients, and curvatures

Figure 2. Interpolated surfaces for (first row: left) data used in simulation experiments (first row: center) true gradient along xx (first row: right) true gradient along yy; (second row: left) true curvature along xx (second row: center) true mixed curvature (second row: right) true curvature along yy; (third row: left) fitted data (third row: center) estimated gradient along xx (third row: right) estimated gradient along yy; (fourth row: left) estimated curvature along xx (fourth row: center) estimated mixed curvature (fourth row: right) estimated curvature along yy.

The true values of gradients are available in closed-form. For example, ∂xμ0(sG)=20cos⁡∥sG∥sx,G∥sG∥\partial_{x}\mu_{0}(s_{G}) = 20 \cos\lVert s_{G} \rVert\frac{s_{x,G}}{\lVert s_{G} \rVert}, where sG=(sx,G,sy,G)s_{G} = (s_{x,G},s_{y,G}) lies on a grid overlaid on the domain of reference (in this case: [−10,10]×[−10,10][-10, 10] \times[-10, 10]). Other gradient and curvature processes are computed similarly by differentiating μ0(s)\mu_{0}(s). The resulting plots are shown in the remaining plots within the first and second rows of Figure 2. We fit the model in eq. (9) to the simulated data using only an intercept, letting ZZ learn the spatial variation.

We fit the model in eq. (9) using a Matérn kernel with ν=52\nu= \frac{5}{2} to the simulated data. This allows for inference on gradients and curvatures. Posterior samples of β\beta, θ\theta are generated from eq. (10) followed by posterior samples for Z(s)Z(s) one-for-one β\beta, θ\theta. Next, we estimate gradients and curvatures using the posterior samples of ϕ\phi, σ2\sigma^2 and Z\mathcal{Z}. Posterior sampling is done one-for-one for samples of ϕ\phi, σ2\sigma^2 and Z\mathcal{Z} using eq. (11). We generated 10,000 posterior samples, with a 5,000 burn-in. The model fit was satisfactory: τ^2=0.384\hat{\tau}^2 = 0.384 (0.145, 1.433) containing the true value of 1, σ^2=344.680\hat{\sigma}^2 = 344.680 (194.292, 687.247) and ϕ^=0.380\hat{\phi} = 0.380 (0.302, 0.489), where the highest posterior density (HPD) intervals are shown in the brackets. The fitted process is shown in the left plot of the third row in Figure 2, while the estimated gradients and curvatures are shown in the remaining plots of the third and fourth rows of Figure 2. We conclude the derivative estimation with Figure 4 which shows further diagnostics for model fit comparing observed with fitted processes, gradients and curvatures.

Diagnostics assessing quality of fit using observed vs. fitted values

Figure 4. Diagnostics assessing quality of fit using observed vs. fitted values for (top: left) response (top: center) ∂x\partial_x (top: right) ∂y\partial_y (bottom: left) ∂xx2\partial^2_{xx} (bottom: center) ∂xy2\partial^2_{xy} (bottom: right) ∂yy2\partial^2_{yy} with 95% credible bands.

The wombling exercise requires a curve. The easiest choice for curves are contours. We choose a closed contour for our purposes. We denote this curve using CC. Posterior samples of Γ(C)\Gamma(C) are generated one-for-one ϕ\phi, σ2\sigma^2 and Z\mathcal{Z} using eq. (12). The left plot of Figure 3 shows CC. It is located in a region where the surface is locally increasing. Line segment level inference for gradient and curvature wombling are shown in the top and bottom plots of the right column of Figure 3 respectively. For the differential processes the percentage of grid locations that had true values contained in their HPD intervals are as follows: 90% for ∂x\partial_x, 92% for ∂y\partial_y, 95% for ∂xx2\partial^2_{xx}, 100% for ∂xy2\partial^2_{xy} and 100% for ∂yy2\partial^2_{yy}. The higher percentages for the second order derivatives is explained by the increased process variance for the processes ∂xx2\partial^2_{xx}, ∂xy2\partial^2_{xy} and ∂yy2\partial^2_{yy} as seen in Theorem 2 resulting in wide HPD intervals. Approximately 96% of the line segments had true wombling measures in their corresponding HPD intervals. For the wombling measure, Γ^(C)=(−108.765,154.565)T\hat{\Gamma}(C) = (-108.765, 154.565)^{\mathsf{T}}, with 95% HPD intervals being (−182.098,−36.290)(-182.098, -36.290) and (69.053,241.664)(69.053, 241.664) respectively, containing the true values Γ(C)=(−131.149,144.010)T\Gamma(C) = (-131.149, 144.010)^{\mathsf{T}}. The curve CC forms a wombling boundary with regard to gradient and curvature. The significance assessments displayed along the curves are based on marginal posterior credible intervals for individual line segments (not point-wise rather, segment wise, each segment connects two consecutive points in the curve). They should therefore be interpreted as segment-specific posterior assessments rather than as a simultaneous test over the entire curve. We compared our approach to Banerjee and Gelfand [5]; Halder et al. [19] using software in Halder and Banerjee [16]—under the same settings for posterior inference our approach produced full posterior inference on wombling measures in 2.06 minutes, while the conventional approach took 15.31 minutes. Figure 5 shows a comparison between popular methods for clustering using density-based scan statistic (DBSCAN) and kk-means, we observe that they only offer inference at the observed location and are unable to offer inference over curves. Our chosen curve partially overlaps the cluster (cluster 1 in DBSCAN and region marked with black points in kk-means).

Plots showing the closed curve and line segment level significance for gradient- and curvature-based wombling measures

Figure 3. Plots showing (left) the closed curve, CC used for wombling (center) line segment level significance for the gradient based wombling measure (right) line segment level significance for the curvature based wombling measure.

Plots showing comparison between approaches

Figure 5. Plots showing comparison between approaches: (left) hot-spot detection via density based clustering, (center) kk-means with three mediods (right) the curve chosen for wombling.

We conclude our experiments by noting that within an ecological setting such curves can either be pre-specified, for example, mountain ranges, rivers or soil and geochemical transitions. Alternatively, they can be chosen by the investigator. Selecting a curve typically requires identifying regions in the spatial domain that sit between high and low abundance, more specifically regions with sharp gradients, which could mark the presence of lurking barriers to the spread of a certain species. We use this principle to select species-specific curves in the next section.

Invasive Species in North America

We apply the Bayesian wombling framework to the spatial occurrence of three invasive species of concern: Adelges tsugae (AT), Lycorma delicatula (LD), and Microstegium vimineum (MV). Retaining all three species is useful because they exhibit markedly different degrees and patterns of spatial variation, allowing us to illustrate the behavior of the proposed framework across contrasting ecological settings.

θ,β\theta, \betaATLDMV
ϕ^\hat{\phi}0.789<br>(0.471, 1.190)2.408<br>(2.048, 2.840)1.835<br>(1.585, 2.110)
τ~2\tilde{\tau}^{2}0.583<br>(0.516, 0.655)0.571<br>(0.455, 0.695)0.508<br>(0.469, 0.555)
σ^2\hat{\sigma}^{2}0.307<br>(0.185, 0.496)2.454<br>(1.796, 3.205)0.551<br>(0.458, 0.666)
β^0\hat{\beta}_{0}0.649<br>(0.571, 0.733)1.972<br>(1.875, 2.064)0.843<br>(0.795, 0.887)
β^bio7\hat{\beta}_{bio7}0.009<br>(0.004, 0.015)-0.079<br>(-0.091, -0.068)0.001<br>(-0.003, 0.005)
β^bio8\hat{\beta}_{bio8}-0.018<br>(-0.028, -0.009)0.132<br>(0.114, 0.151)0.007<br>(0.001, 0.013)
β^bio9\hat{\beta}_{bio9}-0.006<br>(-0.013, 0.002)-0.016<br>(-0.033, 0.001)-0.019<br>(-0.023, -0.015)

Table 1. Parameter and coefficient estimates for different species. The HPD intervals are shown in brackets. Significant effects are marked in bold.

The analysis proceeds in three stages for each species. First, we fit the spatial model in (13) and characterize the residual spatial variation in occurrence after accounting for the bioclimatic covariates. Second, we estimate the local first- and second-order spatial derivative processes. The gradients identify the magnitude and direction of local spatial changes in occurrence, whereas the curvatures characterize changes in those gradients and hence more complex local spatial structure. Third, guided by the estimated spatial surface and its derivatives, we select a candidate curve traversing a region of pronounced spatial change and perform Bayesian wombling. This final step shifts inference from individual locations to the curve itself and evaluates whether the selected curve represents a meaningful spatial boundary.

Because the available data are spatial rather than spatiotemporal, none of these quantities should be interpreted as direct estimates of the temporal movement of an invasion front. Instead, the gradients and wombling boundaries identify sharp spatial transitions in occurrence. When considered together with independent biological and historical information, such transitions may be consistent with invasion boundaries, barriers, corridors, or other features relevant to surveillance and management.

We fit the model

log⁡{Y(s)}=β0+βbio7Xbio7+βbio8Xbio8+βbio9Xbio9+Z(s)+ϵ(s),\log\{Y(\mathbf{s})\} = \beta_{0} + \beta_{bio7}X_{bio7} + \beta_{bio8}X_{bio8} + \beta_{bio9}X_{bio9} + Z(\mathbf{s}) + \epsilon(\mathbf{s}),

where Y(s)Y(\mathbf{s}) denotes the species count at location s\mathbf{s}, Z(s)Z(\mathbf{s}) follows a Gaussian process with a Matérn covariance having ν=5/2\nu= 5/2, and ϵ(s)∼N(0,τ2)\epsilon(\mathbf{s}) \sim N(0,\tau^{2}). The three temperature-related bioclimatic covariates are annual temperature range (bio7), mean temperature of the wettest quarter (bio8), and mean temperature of the driest quarter (bio9).

Table 1 immediately reveals substantial differences among the three species. The proportion of residual variance attributable to the spatial process, σ2/(σ2+τ2)\sigma^{2}/(\sigma^{2}+\tau^{2}), is 34.49% for AT, 81.12% for LD, and 52.03% for MV. Thus, AT exhibits the weakest residual spatial structure, LD the strongest, and MV an intermediate level. These differences provide a useful context for interpreting the derivative and wombling analyses below.

The estimated bioclimatic associations also differ by species. For AT, occurrence is positively associated with bio7 and negatively associated with bio8, whereas for LD these associations are reversed; bio9 is not significant for either species. MV is positively associated with bio8 and negatively associated with bio9, while its association with bio7 is not significant. These species-specific relationships reinforce the need to interpret the subsequent spatial derivative surfaces separately rather than treating the three applications as replications of the same spatial pattern.

Adelges tsugae AT provides the case with the weakest residual spatial signal. Consistent with the variance decomposition above, the estimated spatial effects in Figure 6 are comparatively localized, with significant spatial structure concentrated primarily around the New York–New Jersey–Connecticut region, the Carolinas, and Nova Scotia. The derivative surfaces similarly contain relatively few significant locations. Significant longitudinal and latitudinal gradients occur primarily along portions of the eastern United States, the Carolinas, and Nova Scotia, indicating localized rather than widespread sharp spatial transitions in occurrence. Significant curvature is still more restricted, although the mixed-curvature surface suggests comparatively complex local spatial structure around New York.

Interpolated spatial plots for AT showing estimated spatial effect, gradients, and curvatures

Figure 6. Interpolated spatial plots for AT showing (row 1) estimated spatial effect, estimated gradient ∂x\partial_x along the longitude, estimated gradient ∂y\partial_y along the latitude; (row 2) estimated curvature, ∂xx2\partial_{xx}^{2} along the longitude, estimated mixed curvature, ∂xy2\partial_{xy}^{2} and estimated curvature, ∂yy2\partial_{yy}^{2} along the latitude. Only significant locations and grid points are shown–positive with green and negative with cyan.

For the curve-level analysis, we select a curve traversing the tri-state region, shown in the first row of Figure 9. This region was chosen because the fitted derivative surfaces indicate localized spatial change there. The gradient wombling analysis identifies significant segments primarily along the northern portion of the curve near Connecticut, whereas there is substantially less evidence of significant directional curvature. The estimated total wombling measures are

Spatial plots showing line-segment level inference for gradient and curvature wombling respectively for the three species

Figure 9. Spatial plots showing line-segment level inference for gradient (left column) and curvature (right column) wombling respectively for the three species. Significant segments of curves are highlighted–positive with green and negative with cyan.

Γ^(C)=(0.219,−0.188)T,\widehat{\boldsymbol{\Gamma}}(C) = (0.219,-0.188)^T,

with 95% HPD intervals (−0.002, 0.465) for the gradient measure and (−0.862, 0.506) for the curvature measure. Thus, the curve-level evidence for AT is relatively weak compared with the other species: the gradient measure is close to significance, while the curvature measure clearly includes zero. This result is consistent with the comparatively weak residual spatial structure estimated for AT and illustrates a setting in which wombling identifies localized candidate boundary segments without strong evidence for a boundary over the entire selected curve.

Lycorma delicatula LD presents a sharply contrasting case. It has the largest spatial variance component among the three species, with approximately 81% of residual variation attributable to the spatial process. Correspondingly, Figure 7 shows pronounced spatial structure concentrated around the tri-state region, Pennsylvania, Delaware, and adjoining portions of the eastern United States.

Interpolated spatial plots for LD showing estimated spatial effect, gradients, and curvatures

Figure 7. Interpolated spatial plots for LD showing (row 1) estimated spatial effect, estimated gradient ∂x\partial_x along the longitude, estimated gradient ∂y\partial_y along the latitude; (row 2) estimated curvature, ∂xx\partial_{xx} along the longitude, estimated mixed curvature, ∂xy\partial_{xy} and estimated curvature, ∂yy\partial_{yy} along the latitude. Only significant locations and grid points are shown–positive with green and negative with cyan.

The estimated derivative surfaces reveal stronger rates of spatial change than those observed for AT. Significant longitudinal gradients occur in Ohio, Pennsylvania, and Virginia, indicating pronounced west–east transitions in occurrence, while significant latitudinal gradients are apparent along portions of the coastal region. The magnitudes of the estimated gradients and curvatures are also larger than for the other two species. Although these gradients describe spatial rather than temporal change, their orientation is consistent with independent studies reporting predominantly westward spread and an over-representation of westward dispersal jumps for LD. The agreement therefore provides useful external context for the spatial transitions identified by the model without requiring that the gradients themselves be interpreted as estimates of temporal spread.

We select a curve traversing New York, New Jersey, and Pennsylvania for the wombling analysis (Figure 9, second row). In contrast to AT, a large proportion of the curve consists of segments with significant directional gradients, providing substantially stronger evidence of a spatial boundary. The total wombling measure is

Γ^(C)=(6.398,−6.677)T,\widehat{\boldsymbol{\Gamma}}(C) = (6.398,-6.677)^T,

with 95% HPD intervals (4.346, 8.705) and (−20.543-20.543, 7.758) for the gradient and curvature measures, respectively. The gradient interval excludes zero, establishing the selected curve as a wombling boundary with respect to first-order spatial change. The curvature interval, however, includes zero. Thus, the principal evidence for the LD boundary comes from a persistent directional gradient across the curve rather than from an integrated second-order change.

Microstegium vimineum MV represents an intermediate case in terms of the strength of residual spatial dependence but differs from both insect species in its geographic extent. Significant spatial effects are distributed over a broader portion of the eastern and southern United States (Figure 8), consistent with its wider observed geographic distribution. The derivative surfaces show significant gradients and curvatures along portions of the eastern coast and southern United States. In particular, the first-order surfaces identify pronounced westward and southward spatial changes in occurrence. Unlike AT, where significant derivative behavior is comparatively localized, the MV results indicate spatial transitions distributed across a broader geographic region.

Interpolated spatial plots for MV

Figure 8. Interpolated spatial plots for MV showing (row 1) estimated spatial effect, estimated gradient ∂x\partial_x along the longitude, estimated gradient ∂y\partial_y along the latitude; (row 2) estimated curvature, ∂xx\partial_{xx} along the longitude, estimated mixed curvature, ∂xy\partial_{xy} and estimated curvature, ∂yy\partial_{yy} along the latitude. Only significant locations and grid points are shown–positive with green and negative with cyan.

For wombling, we select an open curve along the eastern region shown in the third row of Figure 9. Most segments along this curve exhibit significant directional changes. The total wombling measure is

Γ^(C)=(3.013,−6.891)T,\widehat{\Gamma}(\mathrm{C}) = (3.013, -6.891)^{T},

with 95% HPD intervals (1.982, 4.133) for the gradient measure and (−11.997-11.997, −1.953-1.953) for the curvature measure. Both intervals exclude zero. Consequently, the selected curve constitutes a wombling boundary with respect to both first- and second-order spatial change. This distinguishes MV from LD, for which the integrated evidence is primarily gradient-based, and from AT, for which the curve-level evidence is considerably weaker.

Taken together, the three applications demonstrate three distinct inferential regimes. AT exhibits comparatively weak spatial dependence and only localized evidence of sharp spatial change; LD exhibits strong spatial dependence and a pronounced gradient-based boundary; and MV exhibits intermediate overall spatial dependence but a geographically broader pattern of spatial change, with evidence for both gradient- and curvature-based wombling. The three species therefore provide complementary rather than redundant illustrations of the proposed framework.

Discussion and Future Work

We have demonstrated the use of Bayesian wombling for detecting regions of rapid spatial change in the occurrence of invasive species. Our framework provides Bayesian inference over curvilinear boundaries and quantifies spatial gradients and curvatures in species occurrence within a hierarchical model that also incorporates bio-climatic covariates. Importantly, because the data analyzed here are spatial rather than spatiotemporal, these gradients characterize spatial variation in occurrence and should not be interpreted as direct estimates of temporal spread or movement of an invasion front. Instead, the detected boundaries identify sharp spatial transitions that may represent invasion boundaries, barriers to spread, corridors, environmental transitions, or other features of the realized species distribution. For some species, the orientation of the identified gradients is consistent with independently observed and modeled patterns of spread, providing biological context for the spatial patterns detected by the model.

There are many avenues for future research. Further enhancements can be made to testing the curve for a wombling boundary via multiple testing. Multiple candidate boundaries can be considered and Γ(C)\Gamma(\mathrm{C}) can be maximized over such boundaries to yield a final choice. We assume that the logarithm of counts is a Gaussian Process which is not ideal considering the data is originally a count. Similar frameworks can be developed for generalized linear models, particularly focusing on zero-inflated models (see, e.g. Halder et al., 2020, 2021, 2023) which provide a more realistic setting for analyzing raw presence/absence counts. Developing Bayesian wombling using generalized linear models (GLMs) would be beneficial in that regards. Extending the methodology to forecast and back-cast the invasion front would support risk management decisions (e.g. decision to eradicate, contain, or monitor a pest), as well as supporting forensic efforts to identify the initial point of establishment for an invasive species. Direct inference on the movement of an invasion front through time would require spatiotemporal occurrence data and lies beyond the spatial framework considered here. This would require extending the spatiotemporal framework in Halder et al. (2026) to GLMs. Additionally, applying this methodology to past invasions could identify whether pest management actions were effective in managing the pest, especially in the case of ’slow-the-spread’ efforts, and provide a decision support tool in managing future pest invasions. Another promising approach is to include altitude in the spread model and perform boundary analysis over a curved geometry (see, e.g. Li et al., 2026). While our approach employs MCMC for model fitting, we recognize opportunities to incorporate alternative, possibly faster, machine learning tools such as Bayesian predictive stacking (see, for e.g., Pan et al., 2025; Zhang et al., 2026) that are implemented in the spStack (see, for e.g., Pan and Banerjee, 2024) for wombling using stacked posterior surfaces. Extending these models to a multivariate setting will allow us to track correlated regions of spread across multiple species. Finally, an open-source software for ecologists can follow the steps laid out in Halder and Banerjee [17] (2026).

Acknowledgements

AH acknowledges funding from the National Science Foundation, Division of Mathematical Sciences (DMS 2515897), the National Institutes of Health, National Institute of Environmental Health Sciences (R56 ES036250, R21 CA305033), and institutional funds at the Dornsife School of Public Health and the Urban Health Collaborative at Drexel University.

References

  1. [1]Abramowitz, M., Stegun, I. A., and Romer, R. H. (1988). Handbook Of Mathematical Functions With Formulas, Graphs, and Mathematical Tables.
  2. [4]Banerjee, S. and Gelfand, A. E. (2003). On smoothness properties of spatial processes. Journal of Multivariate Analysis, 84(1):85–100.DOI
  3. [5]Banerjee, S. and Gelfand, A. E. (2006). Bayesian Wombling: Curvilinear Gradient Assessment Under Spatial Process Models. Journal of the American Statistical Association, 101(476):1487–1501.DOI
  4. [6]Banerjee, S., Gelfand, A. E., and Sirmans, C. F. (2003). Directional Rates of Change Under Spatial Process Models. Journal of the American Statistical Association, 98(464):946–954.
  5. [8]Bivand, R. S. and Wong, D. W. S. (2018). Comparing implementations of global and local indicators of spatial association. TEST, 27(3):716–748.DOI
  6. [9]Bomford, M. (2008). Risk assessment models for establishment of exotic vertebrates in Australia and New Zealand. Invasive Animals Cooperative Research Centre Canberra.
  7. [11]Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. (1996). A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96, pages 226–231, Portland, Oregon. AAAI Press.
  8. [12]Finley, A. O., Datta, A., Cook, B. D., Morton, D. C., Andersen, H. E., and Banerjee, S. (2019). Efficient Algorithms for Bayesian Nearest Neighbor Gaussian processes. Journal of Computational and Graphical Statistics, 28(2):401–414.arxiv.org/abs/1702.00434
  9. [13]Fitzpatrick, M. C., Preisser, E. L., Porter, A., Elkinton, J., Waller, L. A., Carlin, B. P., and Ellison, A. M. (2010). Ecological boundary detection using Bayesian areal wombling. Ecology, 91(12):3448–3455.pubmed.ncbi.nlm.nih.gov/21302814
  10. [16]Halder, A. and Banerjee, S. (2023). spWombling: Bayesian Wombling using Spatial Differential Processes (https://github.com/arh926/spWombling).
  11. [17]Halder, A. and Banerjee, S. (2025). nimblewomble: Bayesian Wombling using ’nimble’.DOI
  12. [19]Halder, A., Banerjee, S., and Dey, D. K. (2024). Bayesian Modeling with Spatial Curvature Processes. Journal of the American Statistical Association, 119(546):1155–1167.DOI
  13. [24]Holway, D. A. (2005). Edge effects of an invasive species across a natural ecological boundary. Biological Conservation, 121(4):561–567.DOI
  14. [25]Horrocks, K. J., Zhang, J., Haye, T., Seehausen, M. L., Maggini, R., Xian, X., Chen, J., Nugnes, F., Collatz, J., Gruber, A., et al. (2024). Biology, impact, management and potential distribution of aromia bungii, a major threat to fruit crops around the world. Journal of Pest Science, 97(4):1725–1747.
  15. [30]Lindenmayer, D., Nix, H., McMahon, J., Hutchinson, M., and Tanton, M. (1991). The conservation of leadbeater’s possum, gymnobelideus leadbeateri (mccoy): a case study of the use of bioclimatic modelling. Journal of Biogeography, pages 371–383.
  16. [33]Panetta, F. and Dodd, J. (1987). Bioclimatic prediction of the potential distribution of skeleton weed chondrilla juncea l. in western australia.
  17. [36]Stein, M. L. (1999). Interpolation of Spatial Data. Springer Series in Statistics. Springer, New York, NY.
  18. [37]Strydom, T. and Poisot, T. (2023). SpatialBoundaries.jl: edge detection using spatial wombling. Ecography, 2023(5):e06609.DOI
  19. [38]Taylor, S., Kumar, L., Reid, N., and Kriticos, D. J. (2012). Climate change and the potential distribution of an invasive shrub, lantana camara l. PloS one, 7(4):e35565.DOI
  20. [39]Václavík, T. and Meentemeyer, R. K. (2009). Invasive species distribution modeling (ism): Are absence data and dispersal constraints needed to predict actual distributions? Ecological modelling, 220(23):3248–3258.DOI

Paper details

vlnc:W_fwrz56jw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents