Title:Modified planar rotator model for efficient and automated prediction of partially sampled data on Cartesian grids

Abstract:We introduce a novel conditional simulation method inspired from statistical physics for the efficient prediction of missing data on partially sampled two-dimensional Cartesian grids. The proposed method is based on the classical planar rotator spin model, which is suitably modified to allow a one-to-one mapping between continuous variables and spin values. The modified planar rotator (MPR) model does not assume a Gaussian probability distribution, while it captures spatial correlations in terms of nearest-neighbor interactions between the spin variables. The only parameter of the MPR model is the temperature, which we propose to estimate by means of an ergodic specific energy matching principle. We then propose an efficient hybrid Monte Carlo algorithm for conducting conditional simulations at the estimated temperature in order to sample the conditional probability distribution at the missing data locations. The relaxation process of the MPR model is fast and the short-range nature of the interactions enables vectorization of the algorithm. Consequently, the MPR model achieves approximately linear scaling of the computational time with the system size. This scaling makes it more suitable for processing big data sets, such as satellite and radar images, than conventional geostatistical approaches. The performance (accuracy and computational speed) of the MPR model is validated with Gaussian synthetic and non-Gaussian real data (atmospheric heat release measurements and Walker lake DEM-based concentrations) via comparisons with common geostatistical gap-filling methods that include nearest neighbors, inverse distance weighted, bilinear, bicubic, and minimum curvature interpolation.