zlf-lll/SMT-Net

Climate Downscaling Dataset (S2S)

A dataset for climate data spatial downscaling from low resolution (16×16) to high resolution (64×64).

Dataset Description

This dataset contains climate variables for training deep learning models to perform statistical downscaling. The data is preprocessed and ready for use with PyTorch-based frameworks.

Data Format

The dataset consists of PyTorch tensor files (.pt) with the following structure:

Key	Shape	Description
LR_input	[C, T, 16, 16]	Low-resolution input (C=7 channels, T=time steps)
HR_target	[C, T, 64, 64]	High-resolution target
HR_topo	[2, 64, 64]	Topographic data (elevation and slope)

Files

File	Size	Description
dict_s2s_train.pt	8.6 GB	Training set
dict_s2s_test.pt	1.8 GB	Test set
dict_s2s_val.pt	873 MB	Validation set
HR_topo.nc	58 KB	Topographic data (NetCDF format)

Input Channels

The 7 input channels in LR_input typically include:

• Temperature (t2m)
• Geopotential height
• U-wind component
• V-wind component
• Relative humidity
• Surface pressure
• Other meteorological variables

Scale Factor

• Input resolution: 16×16
• Output resolution: 64×64
• Scale factor: 4x

Usage

Loading the Data

import torch

# Load training data
train_data = torch.load('dict_s2s_train.pt')

lr_input = train_data['LR_input']    # [C, T, 16, 16]
hr_target = train_data['HR_target']  # [C, T, 64, 64]
hr_topo = train_data['HR_topo']      # [2, 64, 64]

# Transpose to [T, C, H, W] for batch processing
lr_input = lr_input.permute(1, 0, 2, 3)
hr_target = hr_target.permute(1, 0, 2, 3)

print(f"Number of samples: {lr_input.shape[0]}")
print(f"Input shape: {lr_input.shape}")
print(f"Target shape: {hr_target.shape}")

With Hugging Face Datasets

from huggingface_hub import hf_hub_download

# Download specific file
train_path = hf_hub_download(
    repo_id="YOUR_USERNAME/climate-downscaling-s2s",
    filename="dict_s2s_train.pt",
    repo_type="dataset"
)

train_data = torch.load(train_path)

Citation

If you use this dataset, please cite:

@dataset{climate_downscaling_s2s,
  title={Climate Downscaling Dataset for Deep Learning},
  year={2025},
  url={https://huggingface.co/datasets/YOUR_USERNAME/climate-downscaling-s2s}
}

License

This dataset is released under the MIT License.