Enhancing Thunderstorm Potential Area Detection over Northeast India Using INSAT series of Data: A Deep Learning Approach
This study introduces GeoConvNet, a deep learning framework that leverages multi-spectral INSAT satellite data to accurately detect and classify convective cloud structures over Northeast India, demonstrating strong correlation with rainfall and lightning observations while effectively addressing challenges posed by complex terrain through spatial buffering and terrain-based filtering.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Thunderstorms are among nature's most volatile displays, capable of unleashing intense lightning, heavy rain, and damaging winds in a matter of minutes. For meteorologists, the challenge lies in spotting these systems before they fully form, especially when they are just beginning to stir in the atmosphere. Traditional weather models often struggle to predict these rapid changes because the processes happen on such a small scale that computers cannot resolve them quickly enough. Satellites offer a different perspective, watching from above to see the cold, high tops of clouds that signal a storm is brewing. However, looking at a satellite image is not always straightforward; a cold cloud top does not always mean rain is falling directly beneath it, and the complex, mountainous landscapes of places like Northeast India can make the relationship between what is seen from space and what happens on the ground even harder to decipher.
In this context, a team of researchers from the North Eastern Space Applications Centre and the Indian Space Research Organization has developed a new way to watch these storms. They created a computer system called GeoConvNet, which uses a type of artificial intelligence known as deep learning to analyze images from India's INSAT-3D and 3DR weather satellites. Instead of relying on rigid rules to guess where a storm might be, this system learns to recognize the subtle patterns in the light and heat emitted by clouds. It looks at the temperature of the cloud tops and the moisture in the air above them, then sorts every single point in the image into one of four categories: no storm, a weak storm, a strong storm, or a very intense storm. The researchers trained this system using more than 65,000 small slices of satellite images taken during the pre-monsoon season of 2021, teaching it to distinguish between different types of cloud activity.
To see if the system actually worked, the team tested it against real-world data from ten different thunderstorm events that occurred between 2018 and 2024 over Northeast India. They compared the AI's predictions with two independent sources: a global network that tracks lightning strikes and a satellite-based rainfall map called GPM IMERG. The results showed that the system is quite good at identifying the most dangerous parts of a storm. When the AI flagged an area as having a "very deep moist" cloud, it almost always corresponded to a location where lightning was striking or heavy rain was falling. The system achieved a success rate, measured by a standard metric, of between 0.4 and 0.6 under the best conditions, which is a strong performance for such a difficult task. The researchers also found that the system's accuracy improved when they allowed for a small margin of error in location, acknowledging that rain often falls slightly away from the coldest part of the cloud due to wind and the way storms move.
However, the study also revealed where the system faces limits, particularly when the terrain gets steep. In the lowlands and rolling hills, the AI performed consistently well. But as the elevation rose above 2,000 meters, and especially above 3,000 meters, the system's ability to predict rain dropped significantly. This is because high mountains create their own weather patterns; cold clouds can form over peaks without producing rain, or rain can be blocked by the terrain in ways that the satellite cannot easily see from above. The researchers also noted that the system is best at spotting compact, intense storm cores. When a storm is a vast, spreading sheet of clouds with only weak rain, the AI sometimes misses the precipitation because it is trained to look for the intense, deep updrafts that characterize severe thunderstorms.
Ultimately, this work demonstrates that deep learning can be a powerful tool for monitoring severe weather in regions where ground-based radar is scarce. By combining satellite data with an understanding of how mountains and storm structures affect rainfall, the GeoConvNet system offers a reliable way to spot the most dangerous convective clouds. While it is not perfect, especially in high-altitude areas or during weaker storm phases, it provides a physically meaningful way to translate satellite images into actionable information about where thunderstorms are likely to strike. This approach could help improve early warnings for communities in Northeast India, giving them more time to prepare for the sudden and powerful storms that frequently sweep through the region.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.