Visualizing the potential impact of hurricanes on residential areas before they strike is crucial for helping residents prepare and make informed evacuation decisions.
Researchers at MIT have innovated a groundbreaking technique that synthesizes satellite imagery predicting the aftermath of flooding events. This approach merges a generative artificial intelligence model with a physics-based flood simulation, producing realistic aerial images that indicate where flooding is expected to occur depending on storm strength.
As a practical example, the team implemented this method in Houston, creating satellite images that illustrate how various city locations might appear following a storm similar to Hurricane Harvey, which devastated the area in 2017. They evaluated the generated visuals against actual satellite imagery captured after the hurricane, as well as against images produced solely by the AI model without physics integration.
The physics-enhanced method yielded more accurate and lifelike satellite imagery. In contrast, the AI-only approach resulted in unrealistic flood predictions in areas where flooding is physically implausible.
Considered a proof-of-concept, this method showcases how generative AI models can yield reliable, realistic content when combined with physics-based simulations. To extend its application to other regions and forecast flooding from future storms, the approach needs to be trained on a broader suite of satellite imagery to capture regional variations in flood scenarios.
“Our vision is to ultimately use this technology before hurricanes to provide an additional visualization resource for the public,” explains Björn Lütjens, a postdoctoral researcher in MIT’s Department of Earth, Atmospheric, and Planetary Sciences. He led this research project during his doctoral studies in the Department of Aeronautics and Astronautics (AeroAstro). “One significant hurdle is motivating individuals to evacuate when dangers are imminent. This visualization could enhance preparedness.”
To demonstrate the capabilities of their innovative technique, termed the “Earth Intelligence Engine,” the team has made it available for public access.
The research findings are published in the journal IEEE Transactions on Geoscience and Remote Sensing, co-authored by MIT researchers Brandon Leshchinskiy, Aruna Sankaranarayanan, and Dava Newman, who is a professor of AeroAstro and serves as the director of the MIT Media Lab, along with partners from multiple institutions.
Generative AI and Satellite Imagery
This study builds on the team’s previous endeavors to harness generative AI to project future climate scenarios effectively.
“Delivering a hyper-local view of climate impacts appears to be the most effective method to communicate our scientific findings,” states Newman, the senior author of the study. “People resonate with their direct neighborhoods, where their families and friends reside. Presenting localized climate simulations becomes more intuitive and personal.”
The researchers employed a conditional generative adversarial network (GAN)—an advanced machine learning technique that generates realistic images through the competition between two neural networks. The first “generator” network learns from pairs of real data, such as satellite images captured before and after hurricanes. The second “discriminator” network is tasked with distinguishing between these genuine satellite images and those artificially generated by the first network.
Through continuous feedback and improvement between the two networks, the goal is to create synthetic images that are nearly indistinguishable from actual satellite photographs. However, GANs can occasionally produce “hallucinations”—inaccurate features evident within otherwise realistic images.
“These hallucinations can mislead viewers,” warns Lütjens, who has contemplated methods to minimize these distortions, ensuring that generative AI tools can be trusted in sensitive scenarios. “We pondered: how can we utilize generative AI in climate impact predictions where reliable data is paramount?”
A New Approach to Flood Prediction
In this recent effort, the researchers focused on creating trustworthy satellite imagery depicting potential future flooding scenarios that would aid in preparing and possibly evacuating affected individuals.
Typically, decision-makers rely on color-coded maps for understanding potential flooding areas, produced via various physical model outputs beginning with hurricane tracking models. These models integrate wind patterns and strength simulations alongside forecasts of how water may inundate land due to storm surges. A hydraulic model subsequently outlines flood-prone areas based on local flood infrastructure, culminating in a visual flood elevation map of the region.
“The principal question is whether satellite image visualizations can provide a more tangible and emotionally engaging approach than standard color-coded flood maps, while still maintaining trustworthiness,” Lütjens proposes.
The team initially investigated the results of AI-generated satellite imagery reflecting future flooding conditions using GANs trained on real satellite images taken around Houston before and after Hurricane Harvey. Although the synthetic images appeared to look like typical satellite visuals, closer scrutiny revealed flawed predictions, with flooding depicted in areas where it shouldn’t occur, such as elevated terrains.
To mitigate these inaccuracies and enhance the reliability of AI-generated imagery, the researchers integrated the GAN with a physics-based flood model that incorporates essential physical parameters, such as hurricane trajectories, storm surges, and flood patterns. This partnership produced satellite images of Houston that aligned precisely with flood forecasts generated by the model.
“This work demonstrates a tangible approach to merging machine learning with physics for a scenario that’s highly sensitive to risk, allowing us to better analyze the complexities of Earth’s systems and forecast effective actions to safeguard populations,” Newman conveys. “We are eager to see our generative AI tools placed in the hands of local decision-makers, as they could make a significant impact—potentially saving lives.”
This research was supported by various initiatives, including the MIT Portugal Program, the DAF-MIT Artificial Intelligence Accelerator, NASA, and Google Cloud.
Photo credit & article inspired by: Massachusetts Institute of Technology
