Future Floods in Focus: Using AI and Physics to Visualize Disaster Risks

Massachusetts Institute of Technology (MIT) scientists have developed an innovative method that combines generative artificial intelligence (AI) with a physics-based flood model to create realistic satellite images of future flooding events. This tool, called the "Earth Intelligence Engine," generates detailed, birds-eye-view visuals of regions expected to flood based on the strength and trajectory of approaching storms. As a proof of concept, the team tested the method on Houston, simulating flooding scenarios similar to Hurricane Harvey in 2017. The AI-generated images that incorporated the physics-based flood model proved more accurate and reliable than those generated by AI alone, which often included errors like flooding in areas of high elevation.

The method uses a conditional generative adversarial network (GAN) to produce realistic images by training two neural networks: one generates synthetic satellite images, and the other evaluates their accuracy against real images. While traditional GAN models risk "hallucinations," or inaccuracies in their results, the integration of physics-based parameters significantly reduces these errors. This enhanced approach provides a trustworthy visualization tool that could help policymakers and residents better understand potential flooding risks, making it more emotionally engaging and actionable than standard color-coded flood maps.

The research highlights the importance of pairing AI with trustworthy data sources, particularly in risk-sensitive scenarios like natural disasters. By visualizing potential flooding before a hurricane hits, this technology has the potential to enhance public preparedness and decision-making, such as evacuation planning. While further training on diverse satellite images is needed to apply the model to other regions, the method demonstrates how AI, combined with physics, can support community-level decision-making and potentially save lives.