The AI model was trained with data from extreme rainfall events that flooded Assam in June 2022. Credit: Bhubaneswar2009/CC BY 4.0
A weather forecasting model based on artificial intelligence (AI) predicted heavy rainfall in flood-prone Assam more precisely than conventional methods1. The model can forecast different rainfall categories 96 hours in advance with close to 80% accuracy.
Scientists at the Indian Institute of Technology Bhubaneswar, led by Sandeep Pattnaik, integrated AI into traditional dynamical models to improve prediction of heavy rainfall over Assam, in northeast India. The area experienced devastating floods in June 2022 and 2023.
The India Meteorological Department (IMD) and other weather agencies use dynamical models for forecasting. In northeast India these are only around 38% accurate, primarily due to a poor understanding of the complex physical mechanisms behind the region’s heavy rainfall.
The scientists customised U-Net, a deep learning model, to identify rainfall distribution patterns and classify rainfall intensity from light to extremely heavy – the standard IMD categories.
The AI model was given daily accumulated rainfall outputs from high-resolution dynamical weather models at the district scale over Assam. The AI model matched forecasts with observed data and learned to capture biases in the distribution, amount and intensity of rainfall in the traditional model. It then corrected these biases to provide a forecast.
As a case study, the AI model was trained with data from heavy rainfall events in Assam from 15 to 19 June 2022. The AI model accurately warned of heavy rainfall (more than 150mm/day) 72 hours in advance for the Barpeta, Kamrup, Kokrajhar and Nalbari districts.
