Earth, Atmosphere & Environment Sciences
Title : | Cyclone Intensity and Track Prediction using Deep Learning |
Area of research : | Earth, Atmosphere & Environment Sciences |
Focus area : | Cyclone Intensity Prediction |
Principal Investigator : | Dr. Uma Das, Indian Institute Of Information Technology Kalyani, West Bengal |
Timeline Start Year : | 2023 |
Timeline End Year : | 2026 |
Contact info : | umakota@gmail.com |
Details
Executive Summary : | Climate change is a global issue which adversely impacts all life forms. Although widely discussed, a lot remains to be done to withstand and mitigate the effects of climate change, particularly as it is being accelerated due to anthropogenic activities (IPCC, 2022). The immediate and direct outcome in terms of loss of lives and irreparable damage to livelihoods and economy is mostly felt through extreme weather events like Tropical Cyclones (TC) that cannot be evaded. However, prediction of the same is possible, which would allow adequate steps to be taken to minimize its impact and lessen losses. This is called Disaster Risk Reduction (DRR). Several decades of research exist in the domain of TC. State-of-the art predictions of these cataclysmic events are based on NWP models and ensemble means, maintained by all government agencies across the globe. The predictions are continuously updated in the need of the hour, but still suffer errors due to which we are still unable to avoid the losses that come along with it. Prediction of TC intensity and track is a highly important task which should not have any deviations from the actual and must be done with in sufficient time to implement rescue activities. As per the reports published by the IMD, predicting the track of TC is done with an accuracy varying between 70% to 80% for 72-hour lead time. The above drawbacks warrant the need to explore other tools, like in the field of AI and deep learning, for obtaining better TC forecasts. In the recent times, AI and deep learning have seen great success in many arenas and research in its application for TC forecast have also begun and are showing promising results. Accuracies as high as 99% are now being achieved with deep learning models. The automatic feature extraction of these techniques is often questioned for lack of understanding of the details within the ‘black box’, and thus requires in-depth research to establish transparency and trust in its capabilities, particularly for socially relevant problems as in the current proposal. Deep learning models are now being investigated with ‘explainable AI’ techniques and the intermediary layers are investigated to understand how the important features are extracted from the input data. This study thereby proposes to perform in depth investigations of INSAT-3D data using deep learning models to achieve accurate classification of intensity of TC, precursor identification, and 7-day intensity and track prediction. Explainable AI techniques will also be used to understand the inner workings of these models. Trained models will be delivered that can be used as operational regional AI models for TC forecasts. |
Co-PI: | Dr. Oishila Bandyopadhyay, Indian Institute Of Information Technology Kalyani, West Bengal-741235 |
Total Budget (INR): | 26,43,696 |
Organizations involved