BharatBench: Dataset for data-driven weather forecasting over India

TL;DR

This paper introduces BharatBench, a specialized dataset for data-driven medium-range weather forecasting over India, aiming to enhance machine learning applications in this domain.

Contribution

It provides a new, optimized dataset derived from reanalysis data, along with evaluation metrics and baseline models to advance ML-based weather forecasting in India.

Findings

Baseline scores from simple linear regression and deep learning models.

The dataset facilitates benchmarking and development of ML models for Indian weather forecasting.

Addresses current evaluation limitations and outlines future challenges.

Abstract

Advanced weather and climate models use numerical techniques on grided meshes to simulate atmospheric and ocean dynamics, which are computationally expensive. Data-driven approaches are gaining popularity in weather and climate modeling, with a broad scope of applications. Although Machine Learning (ML) has been employed in this domain, significant progress has occurred in the past decade, leading to ML applications that are now competitive with traditional numerical methods. This study presents a user-friendly dataset for data-driven medium-range weather forecasting focused on India. The dataset is derived from IMDAA reanalysis datasets and optimized for ML applications. The study provides clear evaluation metrics and a few baseline scores from simple linear regression techniques and deep learning models. The dataset can be found at https://www.kaggle.com/datasets/maslab/bharatbench, while the codes are available at https://github.com/MASLABnitrkl/BharatBench. We hope this dataset will boost data-driven weather forecasting over India. We also address limitations in the current evaluation process and future challenges in data-driven weather forecasting.