HR-Extreme Dataset (ICLR 2025)

Overview

HR-Extreme is a dataset containing high-resolution feature maps of physical variables for evaluating the performance of cutting-edge models on extreme weather prediction. This dataset focuses on 17 types of extreme weather events spanning the year 2020, based on HRRR data. The dataset is designed to support researchers in weather forecasting, ranging from physical methods to deep learning techniques. Full paper link

Dataset Access

Index File Access

The code for constructing the dataset is available on GitHub:

Dataset Structure

The dataset is organized into the following folder:

202007_202012: Data from July 2020 to December 2020

Each directory contains the dataset in WebDataset format, following Hugging Face recommendations. Every 10 .npz files are aggregated into a single .tar file, named sequentially as i.tar, where i is an integer (e.g., 0001.tar).

Usage

To generate a complete index file, use the script make_datasetall.py with the start date and end date. For example:

python make_datasetall.py 20200101 20200630

To generate the dataset by the index file, use the script make_dataset_by_index_file.py with the index file from the last step and the start date and end date, you should modify the path of the index file in line 171, note here you should make sure fxx=[0] for Herbie as this means you are using the observing data:

python make_dataset_by_index_file.py 20200101 20200630

To make the predictions of NWP model, use the script make_nwp_predictions.py with the index file from the first step and the start date and end date, you should modify the path of the index file in line 101, note here you should make sure fxx=[1] for Herbie as this means the leadtime is 1:

python make_nwp_predictions.py 20200101 20200630

To test the predictions of NWP for example, use the script test_nwp.py with the start date and end date, this file includes how to calculate RMSE with masks:

python test_nwp.py 20200101 20200630

If you don't have space to save NWP predictions, you can combine make_nwp_predictions.py and test_nwp.py to evaluate NWP predictions online!