In a significant advancement for disaster preparedness in the Himalayan region, researchers from the Indian Institute of Technology (IIT) Bhubaneswar have developed a deep learning-based model capable of predicting cloudburst events in Himachal Pradesh and Uttarakhand up to 72 hours in advance. The model claims far greater accuracy than conventional weather models.
Study Published in Springer Nature Journal
A study detailing this breakthrough was published on June 1 in Neural Computing and Applications, a Springer Nature journal. The research analyzed the devastating cloudburst and extreme rainfall events that struck the northwestern Himalayas between August 12 and 16, 2023, resulting in over 140 deaths and triggering flash floods and landslides across the region.
Limitations of Traditional Models
Researchers Sandeep Pattnaik, Hemant Kumar, Dhananjay Trivedi, Omveer Sharma, and Niladri Bihari Puhan noted that traditional numerical weather prediction models often fail to accurately estimate the timing and intensity of short-duration heavy rainfall events over mountainous terrain. To address this, the team designed a “dual-encoder cross-attention fusion transformer” deep learning model that integrates district- and state-level weather patterns for improved forecasting.
Superior Performance Over WRF Models
“With a mean absolute error of less than 9 mm, the suggested model demonstrated superior rainfall estimation, outperforming the ensembles of the Weather Research and Forecasting (WRF) model,” the study stated. The model successfully captured more than six cloudburst events that occurred across Himachal Pradesh and Uttarakhand during the 2023 disaster period.
According to the study, the AI system accurately tracked temporal rainfall variations in key districts such as Mandi, Dehradun, Haridwar, and Pauri Garhwal, while conventional WRF models “barely predict any events.”
District-Level Accuracy
“The DL model successfully captures the rainfall variation between 36 and 48 hours in the case of Mandi, Dehradun, Haridwar, and Pauri Garhwal, whereas the WRF ensemble model is not able to capture this event in any district,” the study noted. The deep learning system achieved heavy rainfall prediction accuracy of 68.4% in Mandi, 67.33% in Dehradun, 54.66% in Haridwar, and 77.7% in Pauri Garhwal.
Implications for Disaster Preparedness
The researchers’ findings could help authorities issue more reliable early warnings in ecologically fragile Himalayan regions that are increasingly vulnerable to extreme weather events linked to climate change. “This is landmark research with direct implications for improving early warning, disaster preparedness, and mitigation,” the paper concluded.
