AI optimizes land use policy, finding hidden keys for better land use

Using global land use and carbon storage data from the past 175 years, researchers at The University of Texas at Austin and Cognizant AI Labs have trained an artificial intelligence system to develop optimal environmental policy solutions that can advance global sustainability initiatives of the United Nations.

The AI tool effectively balances various complex trade-offs to recommend ways of maximizing carbon storage, minimizing economic disruptions and helping improve the environment and people’s everyday lives, according to a paper published today in the journal Environmental Data Science.

The project is among the first applications of the UN-backed Project Resilience, a team of scientists and experts working to tackle global decision-augmentation problems—including ambitious sustainable development goals this decade—through part of a broader effort called AI for Good.

University of Texas at Austin computer scientist Risto Miikkulainen, who helped launch Project Resilience, believes the new AI approach, initially focused on land use, can address an even larger set of challenges, from infectious diseases to food insecurity, with artificial intelligence potentially discovering better solutions than humans.

“There’s always an outcome you want to optimize for, but there’s always a cost,” he said. Amid all of the trade-offs, AI can home in on unexpected pathways to desirable outcomes at various costs, helping leaders selectively pick battles and yield better results.

The secret sauce of the researchers’ system is evolutionary AI. Inspired by the process of natural selection in biological systems, this computational approach starts with a few dozen policy scenarios and predicts how each scenario will impact various economic and environmental costs.

Then, like a digital version of survival of the fittest, policy combinations that don’t balance the trade-offs well are killed off, while the best ones are allowed to reproduce, giving rise to hybrid offspring.

Random mutations are also sprinkled in to help the system explore novel combinations faster. The process then repeats, winnowing poor performers and keeping the best, across hundreds or thousands of scenarios. Like biological evolution, the “generations” of scenarios become ever-more optimized for a set of priorities.

The team used two tools—a recently released set of global land use data going back centuries and a model that correlates land use with carbon fluxes. First, they used this data to train a prediction model to correlate location, land use and carbon over time. Second, they developed a prescription model to help decision makers find optimal land-use strategies to reduce climate change.

The AI system’s recommendations sometimes surprised the team. Although forests are known to be good at storing carbon, the AI prescription model offered a more nuanced approach than converting as much land as possible into forests, regardless of location. For example, it found that replacing crop land with forest is much more effective than replacing range land (which includes deserts and grasslands).

Also, generally, the same land use change at one latitude didn’t yield the same benefits as at another latitude. Ultimately, the system recommended that larger changes should be allocated to locations where it matters more; in essence, it’s more effective to pick your battles.

“You can obviously destroy everything and plant forests, and that would help mitigate climate change,” said Daniel Young, a researcher at Cognizant AI Labs and a Ph.D. student at UT Austin. “But we would have destroyed rare habitats and our food supply and cities. So we need to find a balance and be smart about where we make changes.”

The researchers have turned their model into an interactive tool that decision makers like legislators can use to explore how incentives, such as tax credits for landowners, would be likely to alter land use and reduce carbon.

Land use activities, including agriculture and forestry are estimated to be responsible for nearly a quarter of all human-caused greenhouse gas emissions. Experts believe smart land use changes will be needed to reduce the amount of carbon in the air and thereby slow climate change. According to Miikkulainen and Young, AI offers options that people, businesses and governments otherwise resistant to change may find easier to accept.

An earlier version of the paper was presented at a major machine learning and computational neuroscience conference, NeurIPS, where it won the “Best Pathway to Impact” award at the Climate Change workshop.

The other authors on the paper are Olivier Francon, Elliot Meyerson, Clemens Schwingshackl, Jakob Bieker, Hugo Cunha and Babak Hodjat.