Skip to content

README.md

Latest commit

 

History

History
191 lines (125 loc) · 8.15 KB

README.md

File metadata and controls

191 lines (125 loc) · 8.15 KB

🌍 GeoFAR: Geography‑informed Frequency‑Aware Representations for Climate Downscaling

Official implementation for the ICLR26 paper: GeoFAR

📘 Introduction

Climate super‑resolution aims to reconstruct high‑resolution climate fields from coarse‑resolution inputs. However, deep learning models struggle to recover geography‑dependent high‑frequency structures, such as those induced by complex terrain, coastlines, and regional atmospheric dynamics.

We introduce GeoFAR (Geography‑informed Frequency‑Aware Representation):

  • 🌊 Frequency‑Aware Representation (FAR) to model multi‑frequency spatial structures.
  • 🗺 Geography‑aware implicit representations (Geo‑INR) to incorporate terrain and geographic priors.
  • ⚙️ Compatibility with both deterministic and generative super‑resolution backbones.

GeoFAR improves reconstruction fidelity for climate variables and significantly enhances the recovery of high‑frequency details in climate downscaling tasks.

🧰 Installation

Clone the repository:

git clone https://github.com/eceo-epfl/GeoFAR.git
cd GeoFAR

Create the conda environment:

conda create -n geofar python=3.10
conda activate geofar

Install dependencies:

pip install "climate_learn_ds @ git+https://github.com/eceo-epfl/climate_learn_ds.git"

📦 Dataset

GeoFAR supports CERRA, ERA5, PRISM, and customized datasets for super‑resolution experiments.

🇪🇺 CERRA Downscaling

We evaluate GeoFAR on the CERRA regional climate dataset under two experimental settings.

Single-variable setting

We perform downscaling on 2m temperature:

  • 🌡 2m temperature (t2m)

📦 Download: Hugging Face Dataset

Clone with Git:

git clone https://huggingface.co/datasets/chaseltsui/cerra-t2m

Multi-variable setting [to do]

We jointly downscale multiple surface variables:

  • 🌡 2m temperature (t2m)
  • 🧭 10m wind (10u, 10v)
  • 💧 2m relative humidity (rh2m)
  • 🌍 surface pressure (sp)

Dataset structure

The expected directory structure is:

dataset/
├── cerra/
│   ├── low-res/
│   │   ├── train
│   │   ├── val
│   │   └── test
│   └── high-res/
│       ├── train
│       ├── val
│       └── test

Additional preprocessing scripts can be found in:

tools/

🌐 ERA5 Downscaling

We also evaluate GeoFAR on global climate downscaling tasks using ERA5 data.

Instructions for downloading and preprocessing ERA5 datasets can be found in the climate-learn documentation:

https://climatelearn.readthedocs.io/en/latest/user-guide/tasks_and_datasets.html

🇺🇸 ERA5 → PRISM Downscaling

For global-to-local downscaling, we use ERA5 as the low-resolution input and PRISM as the high-resolution target.

Dataset preparation follows the same procedure described in the climate-learn documentation:

https://climatelearn.readthedocs.io/en/latest/user-guide/tasks_and_datasets.html

🧠 Supported Models

To our knowledge, GeoFAR provides one of the most comprehensive deep-learning toolboxes for climate downscaling, covering classical interpolation baselines, generic image super-resolution backbones, and climate-specific downscaling models. This repository provides a unified benchmarking framework for CNN-, GAN-, Transformer-, diffusion-, and neural-operator-based super-resolution methods.

Model Venue Brief description Link
ResNet CVPR 2016 Residual CNN backbone widely used as a strong SR/downscaling baseline. Paper
U-Net MICCAI 2015 Encoder-decoder architecture with skip connections for dense prediction and super-resolution tasks. Paper
ViT ICLR 2021 Vision Transformer using patch tokens and global self-attention for representation learning. Paper
EDSR CVPRW 2017 Enhanced deep residual network optimized for high-fidelity image super-resolution. Paper
FFL ICCV 2021 The focal frequency loss to mitigate frequency bias of SR. Paper
SRGAN CVPR 2017 GAN-based super-resolution model designed to generate perceptually realistic textures. Paper
SwinIR ICCVW 2021 Transformer-based image restoration network built on Swin Transformer blocks. Paper
SRFormer ICCV 2023 Transformer SR model with permuted self-attention to enlarge receptive fields. Paper
DeepSD KDD 2017 Early deep-learning-based statistical downscaling method for climate data. Paper
FACL NeurIPS 2025 A frequency domain loss for precpitation Nowcasting Paper
SmCL JMLR 2023 Hard-Constrained deep learning for climate downscaling. Paper
DSFNO JMLR 2024 Fourier Neural Operator designed for arbitrary-resolution climate downscaling. Paper
ClimateDiffuse arXiv 2024 Diffusion-based generative model for climate super-resolution. Paper
GeoFAR ICLR 2026 Geography-informed frequency-aware representation framework for climate downscaling. Paper

Notes

  • Generic SR methods: ResNet, U-Net, ViT, EDSR, SRGAN, SwinIR, SRFormer
  • Climate-specific baselines: DeepSD, DSFNO, ClimateDiffuse, FFL, FACL, SmCL
  • GeoFAR variants: The framework supports multiple enhanced backbones such as GeoFAR[U-Net], GeoFAR[ViT], GeoFAR[SRGAN], and GeoFAR[DSFNO].

🚀 Usage

Training

Example command for training on the CERRA 2x single-variable setting:

python experiments/downscaling/cerra_cerra_downscale.py {low-res-path} {high-res-path} vit t2m --ratio=2 --bs=2 --gpu=0

Evaluation

Run evaluation by specifying a checkpoint path:

python experiments/downscaling/cerra_cerra_downscale.py {low-res-path} {high-res-path} vit t2m --ratio=2 --bs=2 --gpu=0 --checkpoint={checkpoint-path}

📑 Citation

If you find GeoFAR useful in your research, please cite:

@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}
}

⭐ If you find this repository useful, please consider giving it a star!