Geospatial Foundation Models to Enable Progress on Sustainable Development Goals

Artificial Intelligence, Earth Observation, Sustainable Development Goals, Foundation Models

The preprint focuses on evaluating and benchmarking existing Foundation Models (FMs) within the established paradigm of applying AI to Earth Observation (EO) data for addressing Sustainable Development Goals (SDGs). It introduces SustainFM, a novel benchmarking framework, and presents empirical results comparing different FMs against traditional deep learning models. While advocating for a “paradigm shift” from model-centric development to impact-driven deployment, the core research activity still falls under normal science, as it operates within the existing AI/EO paradigm, refining methodologies, and evaluating performance within established metrics. The proposed shift, while significant, represents an evolution within the current paradigm, not a complete overthrow of existing assumptions or methodologies. Therefore, it’s classified as normal science with a potential, future trajectory towards a model drift, if the advocacy gains wider adoption, urging the community to re-evaluate existing model-centric practices.