Fast gap-filling of massive data by local-equilibrium conditional simulations on GPU

TL;DR

This paper introduces a GPU-accelerated, local-equilibrium conditional simulation method for efficient gap-filling in massive space-time datasets, improving accuracy and speed over previous approaches.

Contribution

It extends the modified planar rotator method by modeling spatial heterogeneity through variable parameters and applying non-equilibrium simulations for better predictions.

Findings

Significant computational speedup over CPU implementations.

Improved accuracy in gap-filling of large datasets.

Effective modeling of spatial heterogeneity.

Abstract

The ever-growing size of modern space-time data sets, such as those collected by remote sensing, requires new techniques for their efficient and automated processing, including gap-filling of missing values. CUDA-based parallelization on GPU has become a popular way to dramatically increase computational efficiency of various approaches. Recently, we have proposed a computationally efficient and competitive, yet simple spatial prediction approach inspired from statistical physics models, called modified planar rotator (MPR) method. Its GPU implementation allowed additional impressive computational acceleration exceeding two orders of magnitude in comparison with CPU calculations. In the current study we propose a rather general approach to modelling spatial heterogeneity in GPU-implemented spatial prediction methods for two-dimensional gridded data by introducing spatial variability to model parameters. Predictions of unknown values are obtained from non-equilibrium conditional simulations, assuming ``local'' equilibrium conditions. We demonstrate that the proposed method leads to significant improvements in both prediction performance and computational efficiency.