Multi-scale Invariant Fields: Estimation and Prediction

TL;DR

This paper introduces multi-scale invariant fields, characterizes their spectral properties, and demonstrates their application in modeling and predicting precipitation data with MSI properties.

Contribution

It extends the concept of multi-selfsimilar fields to multi-scale invariant fields, providing spectral characterization and practical estimation methods.

Findings

Precipitation data exhibits MSI properties.

MSI fields can be effectively estimated from real data.

Prediction accuracy measured by mean absolute percentage error.

Abstract

Extending the concept of multi-selfsimilar random field we study multi-scale invariant (MSI) fields which have component-wise discrete scale invariant property. Assuming scale parameters as $\lambda_i>1$, $i=1,\ldots,d$ and the parameter space as $(1, \infty)^d$, the first scale rectangle is referred to the rectangle $ (1, \lambda_1)\times \ldots \times (1, \lambda_d)$. Applying certain component-wise geometric sampling of MSI field, the harmonic-like representation and spectral density of the sampled MSI field are characterized. Furthermore, the covariance function and spectral density of the sampled Markov MSI field are presented by the variances and covariances of samples inside first scale rectangle. As an example of MSI field, a two-dimensional simple fractional Brownian sheet (sfBs) is demonstrated. Also real data of the precipitation in some area of Brisbane in Australia for two days (25 and 26 January 2013) are examined. We show that precipitation on this area has MSI property and estimate it as a simple MSI field with stationary increments inside scale intervals. This structure enables us to predict the precipitation in surface and time. We apply the mean absolute percentage error as a measure for the accuracy of the predictions.