Sparse Local Implicit Image Function for sub-km Weather Downscaling

TL;DR

This paper presents SpLIIF, a neural implicit model for high-resolution weather downscaling from sparse data, outperforming existing methods in temperature and wind predictions over Japan.

Contribution

The paper introduces SpLIIF, a novel implicit neural representation approach for weather downscaling that leverages sparse station data and topography, demonstrating superior accuracy.

Findings

Up to 50% improvement in temperature downscaling accuracy

10-20% better wind downscaling compared to baselines

Effective in both in- and out-of-distribution scenarios

Abstract

We introduce SpLIIF to generate implicit neural representations and enable arbitrary downscaling of weather variables. We train a model from sparse weather stations and topography over Japan and evaluate in- and out-of-distribution accuracy predicting temperature and wind, comparing it to both an interpolation baseline and CorrDiff. We find the model to be up to 50% better than both CorrDiff and the baseline at downscaling temperature, and around 10-20% better for wind.