IndiBench: A Machine Learning-Ready Regional Forecasting Dataset over India

IndiBench provides a curated benchmark dataset for machine learning-based regional weather forecasting over the Indian subcontinent. It is built upon the Indian Monsoon Data Assimilation and Analysis (IMDAA) reanalysis dataset, produced under the National Monsoon Mission by NCMRWF, UK Met Office, and IMD.

🌏 Dataset Overview

The original IMDAA dataset is a high-resolution regional reanalysis developed by the National Centre for Medium Range Weather Forecasting (NCMRWF), Ministry of Earth Sciences (MoES), Government of India, in collaboration with the UK Met Office and the India Meteorological Department (IMD). It provides hourly weather data from 1979 to 2020 over the Indian subcontinent at a 0.12° (~12 km) spatial resolution and includes over 57 variables across 63 pressure levels.

However, the raw dataset presents several challenges for machine learning workflows, including download difficulty, lack of standardized splits, and storage in meteorological formats. IndiBench addresses these limitations by offering a clean, ready-to-use subset for ML applications.

📦 Contents

IndiBench includes:

Time range: 2000–2019 (20 years)
Interval: 6-hourly (00, 06, 12, 18 UTC)
Region: 6°N–36.72°N, 66.6°E–97.25°E (~256×256 grid)
Train/Val/Test splits:
- Train: 2000–2017 (~26,500 samples)
- Val: 2018 (~1,500 samples)
- Test: 2019 (~1,500 samples)
Variables: 43 channels (see below)

📑 Variable List

Category	Variables
Single-level	TMP (2m temp), UGRD/VGRD (10m wind), APCP (precip), PRMSL (MSLP), TCDCRO (cloud cover)
Pressure-level	TMP_prl, HGT, UGRD_prl, VGRD_prl, RH — at 50, 250, 500, 600, 700, 850, 925 hPa
Static fields	MTERH (terrain height), LAND (land cover)

💾 Data Formats

IndiBench is released in two formats:

🧪 Zarr Format

Chunked, cloud-native array storage
Compatible with xarray, dask
Suitable for scientific analysis and fast slicing

import xarray as xr
ds = xr.open_zarr("imdaa_bench_incremental.zarr", consolidated=True)

🚀 HDF5 Format

Optimized for ML training
Each .h5 file = one time step with all variables
Pre-split into train/, val/, and test/

import h5py
f = h5py.File("imdaa_bench_h5/train/20010101_00.h5", "r")
print(list(f.keys()))

📜 License and Terms of Use

This dataset is released under the Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

✅ Free for non-commercial, educational, and research use
❌ For commercial use, contact: [email protected]
📧 Send a copy of any publication using this dataset to the same address

🔗 References

For questions, issues, or contributions, please open a Discussion or Issue on the Hugging Face page.