Climate AI Climate Downscaling Geography-aware Learning

GeoFAR: Geography-Informed Frequency-Aware Super-Resolution for Climate Data

Chang Xu†, Gencer Sumbul, Li Mi, Robin Zbinden, Devis Tuia
EPFL, Switzerland · ETH Zurich, Switzerland
ICLR 2026
† Correspondence: chang.xu@epfl.ch
Paper (PDF) Code BibTeX
Overview Method Results Analysis BibTeX

What GeoFAR addresses

Current super-resolution methods struggle to reconstruct geography-dependent high-frequency details (e.g., mountains, coastlines, polar regions) in weather and climate data. GeoFAR introduces frequency-aware representations (FAR) and geography-aware implicit neural representations (Geo-INR) to improve super-resolution fidelity across deterministic and generative SR backbones.

Global + local downscaling Surface + pressure variables Deterministic + generative baselines High-frequency fidelity

Motivation

  • Climate data is dominated by low-frequency components, making high-frequency signals relatively underrepresented during training.
  • This frequency distribution is geography-dependent: plains and oceans are mostly smooth, while mountains, coastlines, and polar regions contain richer high-frequency variability.

Deep neural networks also exhibit frequency bias and tend to fit smooth, majority patterns first, which further amplifies this issue in climate super-resolution.

  • As a result, models often under-reconstruct geography-dependent high-frequency details in terrain-complex regions, leading to inaccurate regional estimates.
  • This effect appears as concentrated high-frequency deficits in the Laplace-filtered prediction bias, especially over polar regions, coastlines, and mountains.
Motivation figure for GeoFAR
Motivation of GeoFAR: mitigating frequency bias and geography-dependent high-frequency loss

Method

GeoFAR provides frequency-aware and geography-informed representations for climate SR, thereby reconstructing fine-grained climate information at high resolution. GeoFAR first employes frequency-aware convolutional kernels to produce frequency-aware representations (FAR). In parallel, Geo-INR maps geographical coordinates to a learnable geography (location+elevation) embedding. GeoFAR then performs pixel-wise feature modulation and refinement before feeding the representation into an SR backbone.

GeoFAR pipeline
Overview of the GeoFAR pipeline, including the frequency-aware representations (FAR) and geography-informed INR (Geo-INR).

FAR: Frequency-aware Representation

Uses DCT basis functions to parameterize frequency-aware convolution kernels (FCK), acting as a filter bank that separates frequency-aware patterns into multiple channels and mitigates low-frequency aggregation.

Geo-INR: Geography-informed INR

Combines band-limited spherical harmonics for location and terrain-differential encoding (elevation + gradients), then projects them with a SIREN-based MLP to produce geography-conditioned embeddings for SR.

Super-resolution Results

Across spatial resolutions

GeoFAR significantly improves baselines on global (ERA5: 2.625° → 5.8125°), global-to-local (ERA5: 2.8125°→PRISM: 0.75°), and local (CERRA: 22km → 11km) downscaling of 2m temperature.

Table 1 results
Comparison of GeoFAR vs. ViT across spatial resolutions.

Across downscaling ratios

GeoFAR is robust to varying downscaling ratios (×2, ×4, ×8) and maintains very low RMSE (under 0.4 unit) for as large as 8x downscaling on the 2m temperature (K) variable from CERRA.

Table 2 and Table 3 results
Comparison of GeoFAR vs. UNet across SR ratios.

Across different atmospheric variables

    GeoFAR remains effective on multi-variable CERRA downscaling (T2m, 10u, 10v, Rh2m, Sp) and ERA5 pressure-level variables (Z500, T850). In particular, the advantage of GeoFAR is particularly pronounced in reducing the bias of surface pressure which is strongly modulated by the topography information.

    Table 2 and Table 3 results
    Comparison of GeoFAR vs. ViT across atmospheric variables on CERRA 2x.

Visualizations

ERA5 Super-Resolution

ERA5 Input 1 ERA5 SR 1
ERA5 Global Downscaling (Input vs SR)
ERA5 Input 2 ERA5 SR 2
ERA5 Global Downscaling (Input vs SR)

CERRA Super-Resolution

CERRA Input 2x CERRA SR 2x
CERRA 2× Downscaling (Input vs SR)
CERRA Input 4x CERRA SR 4x
CERRA 4× Downscaling (Input vs SR)
CERRA Input 8x CERRA SR 8x
CERRA 8× Downscaling (Input vs SR)
CERRA Input 8x CERRA SR 8x
CERRA 2x (Alps, 11km)
CERRA Input 8x CERRA SR 8x
CERRA 4x (Alps, 5.5km)
CERRA Input 8x CERRA SR 8x
CERRA 8x (Alps, 5.5km)

CERRA Multi-Variable Super-Resolution

CERRA 10u Input CERRA 10u SR
10m U-Wind
CERRA Rh2m Input CERRA Rh2m SR
2m Relative Humidity
CERRA Sp Input CERRA Sp SR
Surface Pressure

BibTeX

@inproceedings{
xu2026geofar,
title={GeoFAR: Geography-Informed Frequency-Aware Super-Resolution for Climate Data},
author={Chang Xu and Gencer Sumbul and Li Mi and Robin Zbinden and Devis Tuia},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=0WHpOekph0}
}