How do Self-Supervised Remote Sensing Vision Models Transfer to Downstream Tasks?

The study evaluates six self-supervised geospatial foundation models (GeoFMs) across various downstream tasks, revealing that model performance varies significantly depending on the task and adaptation settings. Key findings include the accessibility of task-relevant information in intermediate transformer layers and the impact of decoder design on segmentation tasks. These insights necessitate more representation-aware evaluation and adaptation strategies for practitioners using GeoFMs in remote sensing applications.