EarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery
When one satellite’s data were altered without any ground change, half of the agents’ correct answers shifted with the alteration.
EarthRecord tests twelve multimodal agents on 238 land-change questions across 56 sites and four decades of satellite records. Agents must handle raw imagery, including cloud masking and cross-sensor harmonization. Leading agents complete most analyses end to end but struggle when answers require mapping from imagery. The controls suggest access to the record shifts answers away from prior knowledge, without reliably grounding them in actual land change.
HF Daily Papers' note
EarthRecord tests twelve multimodal agents on 238 land-change questions across 56 sites and four decades of satellite records. Agents must handle raw imagery, including cloud masking and cross-sensor harmonization. Leading agents complete most analyses end to end but struggle when answers require mapping from imagery. The controls suggest access to the record shifts answers away from prior knowledge, without reliably grounding them in actual land change.
HF Daily Papers' note
score 5