Intelligence Where Data is Born: Vision Transformer-Based On-Orbit Change Detection for Earth Sensing

Intelligence Where Data is Born: Vision Transformer-Based On-Orbit Change Detection for Earth Sensing

Salt Palace Convention Center, Salt Lake City, UT

This paper introduces a model framework for autonomous hyperspectral segmentation that couples a foundation model with a supervised segmentation head to enable compact, high performance on-orbit scene understanding. This approach is designed to deliver end-to-end onboard processing, inference, confidence scoring, and prioritization so the spacecraft performs interpretation at or near image collection time and minimizes downlink usage by transmitting only high-value or high-interest outputs rather than full image streams. A key design element is adaptive intelligence over mission lifetime: low-size, over-the-air partial models updates as small as kb-scale enable post-launch retuning and mission-specific refinement without GB-scale data transfer. By reusing the foundation model weights and uplinking only compact task-specific updates, the system minimizes both training overhead and in-flight model upload requirements. Together, these design choices shift mission value from raw data return to onboard prioritization and selective downlink under constrained communications.