언론인들은 위성과 AI를 결합하여 아마존의 불법 채굴을 폭로하고 있습니다.

In 2018, Joseph Poliszuk fled Venezuela. That year, after exposing corruption in then-President Nicolas Maduro’s administration, he had become the target of lawsuits by wealthy Maduro loyalists. He and several of his colleagues at the independent outlet Armando.info packed up their lives and fled the country under threat of imprisonment. For years, Poliszuk had published stories on Southern Venezuela, which is made up of sparsely populated states that cover large swaths of the Amazon Basin and the Orinoco River Basin. Through field reporting, Poliszuk had exposed illegal gold mines, narcotrafficking operations, and crimes against indigenous groups scattered throughout the region’s rainforests. Now in exile — first working from Colombia, then Mexico — Poliszuk was forced to reimagine how to do his work from thousands of miles away. He began experimenting with satellite-based investigations. Satellite imagery has long helped investigative journalists gather intelligence on conflict zones and track changes in remote landscapes. Now, in a new wave of satellite-based investigations, reporters are leaning on machine learning models to automate parts of this work and scale up their analysis to an unprecedented degree.This innovation is most visible in environmental journalism. Poliszuk is just one in a cohort of South American investigative reporters who have used geospatial data and AI-powered pattern recognition to track illegal mining, large-scale logging operations, and cattle ranching across the Amazon. As illegal gold mining spiked during the COVID-19 pandemic, Poliszuk knew there was a story in documenting the growth of these mines across Venezuela’s rainforests. But manually combing through the satellite images for over 50 million hectares of rainforest wasn’t practical. Poliszuk wondered if he could train a machine learning model to detect the scars of mining pits in these images, as well as the neighboring airstrips that are cut into dense vegetation and used to transport minerals. With financial and editorial support from the Pulitzer Center’s first Rainforest Investigations Network (RIN) fellowship and technical support from the nonprofit Earth Genome, Poliszuk was able to do just that. In January 2022, he co-published his first article using the custom machine learning model in a series in El Pais titled “Corredor Furtivo [Clandestine Corridor].” Poliszuk was able to identify 3,718 gold mining locations in the Venezuelan states of Amazonas and Bolívar. Some of those mines were operating inside protected indigenous lands and Canaima National Park, which is home to Angel Falls, the world’s tallest waterfall. By crosschecking maps identifying mining activity with crime data from Venezuelan authorities, Poliszuk was also able to determine whether the mines were run by Venezuelan syndicates, Colombian guerilla groups, or Brazilian garimpeiro (prospectors). The week after Poliszuk published one of his first stories in the El Pais series, the Venezuelan military announced that it had bombed several illegal airstrips operating in the region. “I have 20 years’ experience covering [illegal mining]…thanks to this technology I can show people the dimension of this phenomenon,” Poliszuk told me. “Thanks to this movement, we have understood that we can track by the air what we cannot prove on foot.” Even for journalists who aren’t working in exile, field reporting in the Amazon comes with a litany of accessibility issues and security risks. Poliszuk said a trip from the Venezuelan capital to one of the mines in the state of Amazonas involves a two-hour flight, a six-hour car ride, a four-hour boat ride, then another four-hour trek through the jungle — often through dangerous territory occupied by armed militia. These same groups often hold monopolies on oil and gas in the region, which can make fuel expensive and difficult to procure. “It’s dangerous. It’s challenging. You cannot go there like you can go from Boston to Washington, or Caracas to Maracaibo,” he said. The same year Poliszuk pitched his project to the Pulitzer Center, Brazilian journalist Hyury Potter incubated a similar investigation with the RIN fellowship. He also used machine models from Earth Genome, which collaborates with many journalists to conduct AI-based environmental and human rights investigations. Potter went on to publish several major investigations in Intercept Brasil that identified hundreds of previously unreported airstrips in the Brazilian Amazon and documented the explosion of illegal gold mining on protected indigenous lands. The New York Times published its own investigation based on the satellite imagery analysis, collaborating with Potter and the Pulitzer Center in the reporting process. “It was like a think tank trying to figure out how to do this work,” said Poliszuk of his time in the RIN fellowship. “It was a very good time to think about a new journalism — another way of doing things.”Based on the strength of these early investigations, the Pulitzer Center decided to build a dedicated platform that uses machine learning to track mining activity across the nine countries that are part of the Amazon Basin. Earth Genome built the interface and contributed the underlying geospatial detection models. The nonprofit advocacy group Amazon Conservation contributed fundraising support and helped develop impact metrics. In 2022, the three partner organizations launched Amazon Mining Watch. “That was the beginning — inspired by the works of Joseph and Hyury, we were able to extrapolate and cover the entire Amazon,” said Gustavo Faleiros, the former director of environmental investigations for the Pulitzer Center. The earliest days of Amazon Mining Watch relied on small, task-specific machine learning models. These models were trained by Earth Genome itself and customized only to identify gold mining sites and airfields in satellite imagery. These days, though, Earth Genome is experimenting with more powerful geospatial foundation models — models pre-trained on huge amounts of data, including satellite imagery, but also radar, land cover, and elevation data. It’s likely these larger models will make geospatial investigations even more accessible to journalists, and not just ones covering the Amazon or illegal mining. “In the same way that people figured out how to do unsupervised training of AI models for text — techniques that grew into large-language models — they have done the same thing in the geospatial data space,” said Edward Boyda, a physicist and co-founder of Earth Genome. “With very little additional input from a user — maybe just a few examples — the model can be effectively tuned to detect a wide variety of objects on the Earth’s surface.” The Pulitzer Center and Earth Genome are now partnering with the nonprofit Code for Africa to bring a similar platform to the African continent. Earlier this month, the organizations announced the launch of Africa Mining Watch. The platform will use geospatial detection models to track mining operations across the tropical bend, a region that includes the Congo Basin, the world’s second-largest tropical rainforest. It’s expected to launch publicly in July. On Earth Day last week, 25 journalists from across Africa took part in a seven-hour virtual mapathon to experiment with the platform and test out its ability to identify mines in their coverage areas. “My hope is that Africa Mining Watch will be a platform that’s not as connected with the gold mining story, but with the strategic minerals story,” Faleiros said, pointing to the cobalt, copper, and coltan mines found across the Congo Basin. Earth Genome is also building its own platform to harness these foundation models for journalists. The tool, Earth Index, allows a reporter, researcher, or policy maker to go into the platform and select a region on the world map. After they select examples of the thing they are interested in identifying — say, an artisanal gold mine in Ghana — the platform highlights other potential gold mines in the region.1In its invitation-only beta stage, Earth Index has been used to investigate illegal logging in Albania, commercial flower farms in Uganda, and palm oil production in Brazil. Boyda says they plan to release Earth Index publicly in late April. “The idea with Earth Index is, instead of giving people the data, give them the tool to make their own data,” said Boyda. “Somebody who’s working in a specific area will know that context better than we ever could. With this tool, they can go and build the data set that they want.” This story has been updated to correct where Earth Genome is based.