Comparative analysis of methods for the similarity assessment of geofields

Abstract

In the general case, the problem of spatial modeling can be formulated as the problem of obtaining estimates of the values of the studied parameters in those locations where no measurements were made. An important class of spatial models are geofields, and the most common representation of geofields are two- and three-dimensional arrays, equipped with metadata, including georeferencing parameters. To solve the problem of determining the similarity of geofields (for example, true and obtained by interpolation of the initial data), in practice, methods are often used that were originally developed for assessing the similarity of images, such as photographs or 3D visualization results. Since these methods were developed to analyze images that differ significantly from geofields, their application to geofields is more of a heuristic than a theory-based ap-proach. In the course of the computational experiments described in the article, the applicability of known methods for quantifying the similarity of geofields was investigated, in particular, for the problem of identifying geofields, which are "snapshots" of a geofield changing over time. For the experiments, 7 sets of geofields were generated, each con-taining five thousand geofields. The results of the computational experiments have shown that the MAE, MSE, QSE, NMI, PSNR, RMSE methods make it possible to quite clearly separate the pairs of geofields that are not related to each other, and the pairs of geofields that are consecutive "snapshots" of a geofield that changes in time.