Earth Observation Foundation Models for region-specific flood segmentation (Papers Track)
Helen Tamura-Wicks (IBM Research); Geoffrey Dawson (IBM Research); Andrew Taylor (Science and Technology Facilities Council); Chris Dearden (Science and Technology Facilities Council); Anne Jones (IBM Research); Paolo Fraccaro (IBM Research)
Abstract
AI foundation models for earth observation are an important tool to inform and adapt to extreme weather events brought on by climate change. Here, we investigate the performance of these models for a region-specific task. We build upon the Prithvi-EO model, which uses optical imagery, and incorporate Synthetic Aperture Radar (SAR) imagery for UK and Ireland by both additional pretraining and directly fine tuning for regional flood segmentation. Incorporating SAR band imagery via either approach improved flood segmentation performance from 0.58 to 0.79 (by approximately 35%), suggesting that EOFMs can relatively easily be tuned to new locations and application-specific satellite bands.