GeoRoPE: Ground-Aware Rotary Adaptation for Remote Sensing Foundation Models
Researchers have developed GeoRoPE, a novel method for adapting remote sensing foundation models (RSFMs) to handle scale mismatches across different sensors and ground sampling distances (GSDs). The approach recalibrates token-level positional interactions by first calibrating geo-coordinates to account for varying ground distances and then adjusting the RoPE frequency based on scene-dependent spatial granularity. GeoRoPE is implemented as a lightweight adapter, preserving the frozen spatial prior of existing RSFMs while introducing geo-aware positional corrections. Experiments show that GeoRoPE enhances cross-resolution robustness and improves scale-sensitive representation learning in RSFMs. AI
IMPACT GeoRoPE offers a method to improve the adaptability of remote sensing foundation models to varying spatial resolutions and granularities, potentially enhancing their performance on diverse downstream tasks.