This article was originally published on Undark.
In a sloping backyard in Vallejo, California, Nicholas Spada adjusted a piece of equipment that looked like a cross between a tripod, a briefcase, and a weather vane. The sleek machine, now positioned near a weathered gazebo and a clawfoot bathtub filled with sun-bleached wood, is meant for inconspicuous sites like this, where it can gather long-term information about local air quality.
Spada, an aerosol scientist and engineer at the University of California, Davis, originally designed the machine for a project based about 16 miles south, in Richmond. For six months, researchers pointed the equipment—which includes a camera, an air sensor, a weather station, and an artificial intelligence processor—at railroad tracks transporting coal through the city, and trained an AI model to recognize trains and record how they affected air quality. Now Spada is scouting potential locations for the sensors in Vallejo, where he collaborates with residents concerned about what’s in their air.
The project in Richmond was Spada’s first using AI. The corresponding paper, which published in March 2023, arrived amid proliferating interest—and concern—about AI. Technology leaders have expressed concern about AI’s potential to displace human intelligence; critics have questioned the technology’s potential bias and harvest of public data; and numerous studies and articles have pointed to the significant energy use and greenhouse gas emissions associated with processing data for its algorithms.
But as concern has sharpened, so has scientific interest in AI’s potential uses—including in environmental monitoring. From 2017 to 2021, the number of studies published each year on AI and air pollution jumped from 50 to 505, which an analysis published in the journal Frontiers in Public Health attributed, in part, to an uptick of AI in more scientific fields. And according to researchers like Spada, applying AI tools could empower locals who have long experienced pollution, but had little data to explicitly prove its direct source.
In Richmond, deep learning technology—a type of machine learning—allowed scientists to identify and record trains remotely and around the clock, rather than relying on the traditional method of in-person observations. The team’s data showed that, as they passed, trains full of coal traveling through the city significantly increased ambient PM2.5, a type of particulate matter that has been linked to respiratory and cardiovascular diseases, along with early death. Even short-term exposure to PM2.5 can harm health.
The paper’s authors were initially unsure how well the technology would suit their work. “I’m not an AI fan,” said Bart Ostro, an environmental epidemiologist at UC Davis and the lead author of the paper. “But this thing worked amazingly well, and we couldn’t have done it without it.”
Ostro said the team’s results could help answer a question few researchers have examined: How do coal facilities, and the trains that travel between them, impact air in urban areas?
That question is particularly relevant in nearby Oakland, which has debated a proposed coal export terminal for nearly a decade. After Oakland passed a resolution to stop the project in 2016, a judge ruled that the city hadn’t adequately proved that shipping coal would significantly endanger public health. Ostro and Spada designed their research in part to provide data relevant to the development.
“Now we have a study that provides us with new evidence,” said Lora Jo Foo, a longtime Bay Area activist and a member of No Coal in Oakland, a grassroots volunteer group organized to oppose the terminal project.
The research techniques could also prove useful far beyond the Bay Area. The AI-based methodology, Foo said, can be adapted by other communities looking to better understand local pollution.
“That’s pretty earth shattering,” she said.
Across the United States, around 70 percent of coal travels by rail, transiting from dozens of mines to power plants and shipping terminals. Last year, the U.S.—which holds the world’s largest supplies of coal—used about 513 million tons of coal and exported about another 85 million tons to countries including India and the Netherlands.
Before coal is burned in the U.S. or shipped overseas, it travels in open-top trains, which can release billowing dust in high winds and as the trains speed along the tracks. In the past, when scientists have researched how much dust these coal trains release, their research has relied on humans to identify train passings, before matching it with data collected by air sensors. About a decade ago, as domestically-produced natural gas put pressure on U.S. coal facilities, fossil fuel and shipping companies proposed a handful of export terminals in Oregon and Washington to ship coal mined in Wyoming and Montana to other countries. Community opposition was swift. Dan Jaffe, an atmospheric scientist at the University of Washington, set out to determine the implications for air quality.
In two published studies, Jaffe recorded trains in Seattle and the rural Columbia River Gorge with motion sensing cameras, identified coal trains, and matched them with air data. The research suggested that coal dust released from trains increased particulate matter exposure in the gorge, an area that hugs the boundary of Oregon and Washington. The dust, combined with diesel pollution, also affected air quality in urban Seattle. (Ultimately, none of the planned terminals were built. Jaffe said he’d like to think his research played at least some role in those decisions.)
Studies at other export locations, notably in Australia and Canada, also used visual identification and showed increases in particulate matter related to coal trains.
Wherever there are coal facilities, there will be communities nearby organizing to express their concern about the associated pollution, according to James Whelan, a former strategist at Climate Action Network Australia who contributed to research there. “Generally, what follows is some degree of scientific investigation, some mitigation measures,” he said. “But it seems it’s very rarely adequate.”
Some experts say that the AI revolution has the potential to make scientific results significantly more robust. Scientists have long used algorithms and advanced computation for research. But advancements in data processing and computer vision have made AI tools more accessible.
With AI, “all knowledge management becomes immensely more powerful and efficient and effective,” said Luciano Floridi, a philosopher who directs the Digital Ethics Center at Yale University.
The technique used in Richmond could also help monitor other sources of pollution that have historically been difficult to track. Vallejo, a waterfront city about 30 miles northeast of San Francisco, has five oil refineries and a shipyard within a 20 mile radius, making it hard to discern a pollutant’s origin. Some residents hope more data may help attract regulatory attention where their own concerns have not.
“We have to have data first, before we can do anything,” said Ken Szutu, a retired computer engineer and a founding member of the Vallejo Citizen Air Monitoring Network, sitting next to Spada at a downtown cafe. “Environmental justice—from my point of view, monitoring is the foundation.”
Air scientists like Spada have relied on residents to assist with that monitoring—opening up backyards for their equipment, suggesting sites that may be effective locations, and, in Richmond, even calling in tips when coal cars sat at the nearby train holding yard.
Spada and Ostro didn’t originally envision using AI in Richmond. They planned their study around ordinary, motion-detecting security cameras with humans—some community volunteers—manually identifying whether recordings showed a train and what cargo they carried, a process that likely would have taken as much time as data collection, Spada said. But the camera system wasn’t sensitive enough to pick up all the trains, and the data they did gather was too voluminous and overloaded their server. After a couple of months, the researchers pivoted. Spada had noticed the AI hype and decided to try it out.
The team planted new cameras and programmed them to take a photo each minute. After months of collecting enough images of the tracks, UC Davis students categorized them into groups—train or no train, day or night—using Playstation controllers. The team created software designed to play like a video game, which sped up the process, Spada said, by allowing the students to filter through more images than if they simply used a mouse or trackpad to click through pictures on a computer. The team used those photos and open-source image classifier files from Google to train the model and the custom camera system to sense and record trains passing. Then the team identified the type of trains in the captured recordings (a task that would have required more complex and expensive computing power if done with AI) and matched the information with live air and weather measurements.
The process was a departure from traditional environmental monitoring. “When I was a student, I would sit on a street corner and count how many trucks went by,” said Spada.
Employing AI was a “game changer” Spada added. The previous three studies on North American coal trains combined gathered data on less than 1,000 trains. The Davis researchers were able to collect data from more than 2,800.
In early July 2023, lawyers for the city of Oakland and the proposed developer of the city’s coal terminal presented opening arguments in a trial regarding the project’s future. Oakland has alleged that the project’s developer missed deadlines, violating the terms of the lease agreement. The developer has said any delays are due to the city throwing up obstructions.
If Oakland prevails, it will have finally defeated the terminal. But if the city loses, it can still pursue other routes to stop the project, including demonstrating that it represents a substantial public health risk. The city cited that risk—particularly related to air pollution—when it passed a 2016 resolution to keep the development from proceeding. But in 2018, a judge said the city hadn’t shown enough evidence to support its conclusion. The ruling said Jaffe’s research didn’t apply to the city because the results were specific to the study location and the composition of the coal being shipped there was unlikely to be the same because Oakland is slated to receive coal from Utah. The judge also said the city ignored the terminal developer’s plans to require companies to use rail car covers to reduce coal dust. (Such covers are rare in the U.S., where companies instead coat coal in a sticky liquid meant to tamp down dust.)
Dhawal Majithia, a former student of Spada’s, helped develop code that runs the equipment used to capture and recognize images of trains while monitoring air quality. The equipment—which includes a camera, a weather station, and an artificial intelligence processor—was tested on a model train set before being deployed in the field. Visual: Courtesy of Bart Ostro/UC Davis
Environmental groups point to research from scientists like Spada and Ostro as evidence that more regulation is needed, and some believe AI techniques could help buttress lawmaking efforts.
Despite its potential for research, AI may also cause its own environmental damage. A 2018 analysis from OpenAI, the company behind the buzzy bot ChatGPT, showed that computations used for deep learning were doubling every 3.4 months, growing by more than 300,000 times since 2012. Processing large quantities of data requires significant energy. In 2019, based on new research from the University of Massachusetts, Amherst, headlines warned that training one AI language processing model releases emissions equivalent to the manufacture and use of five gas-powered cars over their entire lifetime.
Researchers are only beginning to weigh an algorithm’s potential benefits with its environmental impacts. Floridi at Yale, who said AI is underutilized, was quick to note that the “amazing technology” can also be overused. “It is a great tool, but it comes with a cost,” he said. “The question becomes, is the tradeoff good enough?”
A team at the University of Cambridge in the U.K. and La Trobe University in Australia has devised a way to quantify that tradeoff. Their Green Algorithms project allows researchers to plug in an algorithm’s properties, like run time and location. Loïc Lannelongue, a computational biologist who helped build the tool, told Undark that scientists are trained to avoid wasting limited financial resources in their research, and believes environmental costs could be considered similarly. He proposed requiring environmental disclosures in research papers much like those required for ethics.
In response to a query from Undark, Spada said he did not consider potential environmental downsides to using AI in Richmond, but he thinks the project’s small scale would mean the energy used to run the model, and its associated emissions, would be relatively insignificant.
For residents experiencing pollution, though, the outcome of the work could be consequential. Some activists in the Bay Area are hopeful that the study will serve as a model for the many communities where coal trains travel.
Other communities are already weighing the potential of AI. In Baltimore, Christopher Heaney, an environmental epidemiologist at Johns Hopkins University, has collaborated with residents in the waterfront neighborhood of Curtis Bay, which is home to numerous industrial facilities including a coal terminal. Heaney worked with residents to install air monitors after a 2021 explosion at a coal silo, and is considering using AI for “high dimensional data reduction and processing” that could help the community attribute pollutants to specific sources.
Szutu’s citizen air monitoring group also began installing air sensors after an acute event; in 2016 an oil spill at a nearby refinery sent fumes wafting towards Vallejo, prompting a shelter-in-place order and sending more than 100 people to the hospital. Szutu said he tried to work with local air regulators to set up monitors, but after the procedures proved slow, decided to reach out to the Air Quality Research Center at UC Davis, where Spada works. The two have been working together since.
On Spada’s recent visit to Vallejo, he and an undergraduate student met Szutu to scout potential monitoring locations. In the backyard, after Spada demonstrated how the equipment worked by aiming it at an adjacent shipyard, the team deconstructed the setup and lugged it back to Spada’s Prius. As Spada opened the trunk, a neighbor, leaning against a car in his driveway, recognized the group.
“How’s the air?” he called out.
Emma Foehringer Merchant is a journalist who covers climate change, energy, and the environment. Her work has appeared in the Boston Globe Magazine, Inside Climate News, Greentech Media, Grist, and other outlets.
This article was originally published on Undark. Read the original article.
The post Scientists are using AI to track coal train dust appeared first on Popular Science.
Articles may contain affiliate links which enable us to share in the revenue of any purchases made.
from | Popular Science https://ift.tt/r60YGKV
0 Comments