New aI Tool Generates Realistic Satellite Images Of Future Flooding

Visualizing the possible impacts of a typhoon on individuals’s homes before it strikes can assist citizens prepare and decide whether to evacuate.

MIT researchers have developed a technique that creates from the future to illustrate how an area would care for a prospective flooding event. The approach combines a generative expert system design with a physics-based flood design to create practical, birds-eye-view pictures of a region, revealing where flooding is likely to happen given the strength of an oncoming storm.

As a test case, the team applied the approach to Houston and created satellite images depicting what specific locations around the city would look like after a storm similar to Hurricane Harvey, which hit the area in 2017. The group compared these created images with real satellite images taken of the same areas after Harvey hit. They likewise compared AI-generated images that did not consist of a physics-based flood design.

The team’s physics-reinforced method produced satellite pictures of future flooding that were more reasonable and accurate. The AI-only approach, in contrast, generated images of flooding in locations where flooding is not physically possible.

The group’s method is a proof-of-concept, meant to demonstrate a case in which generative AI designs can generate practical, credible material when coupled with a physics-based design. In order to apply the approach to other areas to portray flooding from future storms, it will need to be trained on lots of more satellite images to learn how flooding would look in other areas.

„The idea is: One day, we might use this before a cyclone, where it offers an extra visualization layer for the public,“ states Björn Lütjens, a postdoc in MIT’s Department of Earth, Atmospheric and Planetary Sciences, who led the research study while he was a doctoral trainee in MIT’s Department of Aeronautics and Astronautics (AeroAstro). „One of the biggest challenges is encouraging individuals to leave when they are at risk. Maybe this might be another visualization to help increase that preparedness.“

To illustrate the potential of the new technique, which they have dubbed the „Earth Intelligence Engine,“ the group has made it readily available as an online resource for others to attempt.

The researchers report their results today in the journal IEEE Transactions on Geoscience and Remote Sensing. The study’s MIT co-authors consist of Brandon Leshchinskiy; Aruna Sankaranarayanan; and Dava Newman, professor of AeroAstro and director of the MIT Media Lab; together with partners from multiple institutions.

Generative adversarial images

The brand-new study is an extension of the group’s efforts to apply generative AI tools to imagine future climate circumstances.

„Providing a hyper-local viewpoint of environment seems to be the most effective method to interact our scientific outcomes,“ says Newman, the study’s senior author. „People connect to their own postal code, their regional environment where their friends and family live. Providing local climate simulations ends up being user-friendly, personal, and relatable.“

For this study, the authors utilize a conditional generative adversarial network, or GAN, a kind of artificial intelligence method that can generate practical images utilizing two contending, or „adversarial,“ neural networks. The first „generator“ network is trained on pairs of genuine information, such as satellite images before and after a typhoon. The 2nd „discriminator“ network is then trained to compare the genuine satellite imagery and the one synthesized by the first network.

Each network automatically enhances its efficiency based upon feedback from the other network. The idea, then, is that such an adversarial push and pull need to eventually produce synthetic images that are equivalent from the genuine thing. Nevertheless, GANs can still produce „hallucinations,“ or factually inaccurate features in an otherwise reasonable image that shouldn’t be there.

„Hallucinations can misinform viewers,“ says Lütjens, who began to wonder whether such hallucinations might be avoided, such that generative AI tools can be trusted to help inform individuals, particularly in risk-sensitive circumstances. „We were believing: How can we utilize these generative AI models in a climate-impact setting, where having trusted data sources is so crucial?“

Flood hallucinations

In their new work, the researchers considered a risk-sensitive circumstance in which generative AI is entrusted with creating satellite pictures of future flooding that could be reliable adequate to notify decisions of how to prepare and potentially evacuate individuals out of harm’s method.

Typically, policymakers can get an idea of where flooding might take place based upon visualizations in the kind of color-coded maps. These maps are the last product of a pipeline of physical models that normally starts with a hurricane track design, which then feeds into a wind model that replicates the pattern and strength of winds over a regional area. This is integrated with a flood or storm surge design that forecasts how wind may press any nearby body of water onto land. A hydraulic design then maps out where flooding will take place based on the local flood facilities and produces a visual, color-coded map of flood elevations over a specific area.

„The question is: Can visualizations of satellite imagery add another level to this, that is a bit more tangible and emotionally interesting than a color-coded map of reds, yellows, and blues, while still being trustworthy?“ Lütjens states.

The team first tested how generative AI alone would produce satellite images of future flooding. They trained a GAN on real satellite images taken by satellites as they passed over Houston before and after Hurricane Harvey. When they entrusted the generator to produce new flood pictures of the exact same regions, they found that the images looked like typical satellite images, however a closer appearance exposed hallucinations in some images, in the type of floods where flooding must not be possible (for example, in locations at higher elevation).

To reduce hallucinations and increase the reliability of the AI-generated images, the team combined the GAN with a physics-based flood model that integrates real, physical criteria and phenomena, such as an approaching hurricane’s trajectory, storm surge, and flood patterns. With this physics-reinforced approach, the team produced satellite images around Houston that illustrate the very same flood level, pixel by pixel, as anticipated by the flood model.