A Generalized Convolution Model and Estimation for Non-stationary Random Functions

TL;DR

This paper introduces a flexible convolution-based model for non-stationary random functions, enabling better spatial dependence modeling and prediction in non-stationary environments, with practical inference and simulation methods demonstrated on real data.

Contribution

It proposes a novel convolution model with spatially varying weights for non-stationary covariance functions and develops a parameter inference method using local variogram estimation.

Findings

Outperforms stationary models in soil and rainfall data predictions.

Provides a flexible class of non-stationary covariance functions.

Enables conditional simulation within the non-stationary framework.

Abstract

Standard geostatistical models assume second order stationarity of the underlying Random Function. In some instances, there is little reason to expect the spatial dependence structure to be stationary over the whole region of interest. In this paper, we introduce a new model for second order non-stationary Random Functions as a convolution of an orthogonal random measure with a spatially varying random weighting function. This new model is a generalization of the common convolution model where a non-random weighting function is used. The resulting class of non-stationary covariance functions is very general, flexible and allows to retrieve classes of closed-form non-stationary covariance functions known from the literature, for a suitable choices of the random weighting functions family. Under the framework of a single realization and local stationarity, we develop parameter inference procedure of these explicit classes of non-stationary covariance functions. From a local variogram non-parametric kernel estimator, a weighted local least-squares approach in combination with kernel smoothing method is developed to estimate the parameters. Performances are assessed on two real datasets: soil and rainfall data. It is shown in particular that the proposed approach outperforms the stationary one, according to several criteria. Beyond the spatial predictions, we also show how conditional simulations can be carried out in this non-stationary framework.