Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sparse Local Implicit Image Function for sub-km Weather Downscaling

Yago del Valle Inclan Redondo, Enrique Arriaga-Varela, Dmitry Lyamzin, Pablo Cervantes, Tiago Ramalho

TL;DR

SpLIIF tackles downscaling weather fields from sparse observations by learning a continuous implicit representation that fuses irregular station data with high-resolution topography. The architecture merges sparse inputs with dense gridded data and a topography-conditioned latent, refined by an Enhanced Deep Super-Resolution decoder to produce outputs at arbitrary coordinates. Compared with a simple interpolation baseline and CorrDiff diffusion-based downscaling, SpLIIF achieves up to about 50% RMSE reduction for temperature and ~20% for wind, with especially large gains in complex terrain and under out-of-distribution conditions. These results demonstrate the potential of implicit representations, combined with high-resolution topography, to generate physically plausible, site-specific weather fields from sparse data, with practical relevance for climate adaptation and resilience, and point to future work extending to precipitation and physics-informed hybrids.

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.

Sparse Local Implicit Image Function for sub-km Weather Downscaling

TL;DR

SpLIIF tackles downscaling weather fields from sparse observations by learning a continuous implicit representation that fuses irregular station data with high-resolution topography. The architecture merges sparse inputs with dense gridded data and a topography-conditioned latent, refined by an Enhanced Deep Super-Resolution decoder to produce outputs at arbitrary coordinates. Compared with a simple interpolation baseline and CorrDiff diffusion-based downscaling, SpLIIF achieves up to about 50% RMSE reduction for temperature and ~20% for wind, with especially large gains in complex terrain and under out-of-distribution conditions. These results demonstrate the potential of implicit representations, combined with high-resolution topography, to generate physically plausible, site-specific weather fields from sparse data, with practical relevance for climate adaptation and resilience, and point to future work extending to precipitation and physics-informed hybrids.

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.
Paper Structure (6 sections, 2 figures)

This paper contains 6 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Results
  4. SpLIIF vs Baseline
  5. SpLIIF vs CorrDiff
  6. Conclusion and future work

Figures (2)

  • Figure 1: Percentage improvement in RMSE of SpLIIF over the baseline by altitude and number of input stations for (a) temperature, (b) wind speed, and (c) wind angle. Shaded regions indicate the standard deviation on the mean across the different time slices and spatial patches.
  • Figure 2: (a-c) Probability density functions (PDFs) of absolute errors for downscaled predictions CorrDiff, SpLIIF using sparse JMA station data as input (SpLIIF-JMA), and SpLIIF using ERA5 as input (SpLIIF-ERA5). Error distributions are shown for (a) temperature (°C), (b) wind speed (m/s), and (c) wind direction (°). (d) Example inference during a cold front. First column shows topography, and the next three columns respectively show the temperature and wind from ERA5, SpLIIF-ERA5 and CorrDiff. Bolded arrows are a guide-to-the-eye for wind along a valley.