Devastating floods that claim lives and displace thousands serve as a grim reminder of how unpredictable and destructive water can be. This vulnerability prompted researchers at the Indian Institute of Technology (IIT) Bombay to develop a sophisticated double-barreled artificial intelligence (AI) system designed to predict not just where the next big flood will hit, but exactly how deep the water will rise.
By combining satellite radar data with advanced machine learning, the team has created a high-resolution mapping system that identifies flood-prone zones with over 93% accuracy. This has been prepared to encompass an area of 55,000 square kilometres stretching from Tadri in the Uttara Kannada district of Karnataka to Kanyakumari along the coast of the Western Ghats in South India. This new system promises a powerful tool to shield millions living in India's most vulnerable coastal regions.
Traditionally, flood prediction relied on extensive historical rainfall data and physical sensors. The IIT Bombay team, researcher Dr. Kashish Sadhwani and Professor T.I. Eldho, turned towards pattern recognition, analysing several key conditioning factors.
Interestingly, the study found that surface runoff was a more critical predictor than the volume of rainfall itself.
As Dr. Sadhwani explains, “While rainfall is the primary driver of flood events, it does not directly translate into inundation at a given location. Surface runoff represents the integrated hydrological response of the landscape, capturing the combined effects of rainfall intensity, soil moisture, land use, infiltration capacity, and drainage characteristics.”
To process this data, the team used a two-step process. First, a classification model identifies if an area is at risk. Then, a regression model estimates the depth of the water estimated to accumulate, creating a continuous map of potential inundation. To train this model, the team utilised the European Space Agency’s Sentinel-1 Synthetic Aperture Radar (SAR) satellite images, which can effectively penetrate through monsoon clouds and record observations. By comparing images from before and during past floods, the model learned to recognise the dark shades in the images that indicate standing water.
Ultimately, the model provides high-resolution mapping down to a 30-meter grid, but it currently operates with an error margin (RMSE) of approximately 0.99 meters. While a variation of about one meter is significant in urban planning, Dr. Sadhwani notes that the system's current value lies in its breadth and speed.
“The model is designed for rapid, regional-scale flood assessment, offering high computational efficiency and the ability to generate flood extent and depth information quickly over large areas,” Dr. Sadhwani says. “This makes it particularly valuable for early-stage planning, prioritisation of vulnerable zones, and emergency response support.”
The system also currently focuses on terrains with a slope of less than 7%. It is a deliberate methodological choice as stated in the study, “flood depth calculations were restricted to areas where the slope was less than 7% to ensure accurate flood inundation mapping with SAR images, considering the potential for water movement during image capture”.
Additionally, in steeper terrains, radar signals are affected by geometric distortions such as shadow and layover, and water movement during acquisition can lead to inaccuracies in delineating flood extent and depth. Therefore, applying the slope threshold helps ensure that the derived flood depths are reliable and physically consistent.
For Kerala and Karnataka, this tool is a potential game-changer. Areas where clayey soils trap water and low-lying coastal plains dominate, having a 30-meter resolution map, can enable local authorities to see exactly which hospital, school, or road is most likely to be submerged.
Dr. Sadhwani emphasises that the framework can “identify flood-prone areas, guiding urban planning and land use management,” and plays a “critical role in disaster preparedness and response by enabling authorities to allocate resources effectively and prioritise vulnerable regions for evacuation and relief efforts.”
While the current study focused on the Southern part of the Western Coast of South India, the researchers believe the framework is ready to be scaled up to complex urban hubs such as Mumbai or the East Coast. However, moving into these areas will require accounting for new variables and may need recalibrating and retraining the model.
“Coastal environments introduce additional complexities, such as tidal fluctuations, storm surges, sea-level variations, and drainage backflow effects,” Dr. Sadhwani explains. “The methodology can be effectively adapted by incorporating these coastal-specific parameters into the existing framework.”
As climate change increases the frequency of extreme weather, this AI system offers more than just a forecast; it helps disaster management teams to prepare and plan accordingly to reduce flood-related losses and improve regional resilience.
