Embedding-Only Uplink for Onboard Retrieval Under Shift in Remote Sensing

A new research paper by Sangcheol Sim, accepted at the ICLR 2026 ML4RS workshop, presents a breakthrough method for satellite-based AI that radically reduces data transmission needs. The core innovation is an 'embedding-only uplink' pipeline. Instead of a satellite struggling to downlink gigabytes of raw image pixels to Earth for analysis, a ground station first sends up only compact, pre-computed AI embeddings and metadata. These embeddings, generated using models like OlmoEarth, act as dense numerical representations of visual concepts. Once onboard, the satellite performs a lightweight vector similarity search against a small database of these embeddings to instantly triage new image captures for specific hazards, all while using less than 1 kilobyte of telemetry per query.

The study rigorously tested this approach against 'remote-sensing shift'—real-world challenges like analyzing scenes from different times (pre/post-disaster), different disaster locations, and across 15 distinct geographic cloud sites. On a scaled benchmark of 27 Sentinel-2 satellite scenes, the system demonstrated that the compact embeddings are the key enabler, but the optimal decision logic is task-dependent. For cloud classification, a k-nearest neighbors (kNN) retrieval head performed best (0.92 vs. 0.91 F1-score). For temporal change detection (like spotting new damage), using simple class centroids was vastly superior (0.85 vs. 0.48 F1-score). This finding is crucial: once the sub-1KB embeddings are onboard, the system can switch between these optimal 'heads' for different tasks without any additional costly communication from Earth.

This work fundamentally rearchitects how we think about edge AI in space. It moves the heavy lifting of generating powerful visual embeddings to ground-based infrastructure, where compute and power are abundant. The satellite itself becomes a smart, queryable sensor that asks 'does this look like a hazard I know?' rather than a dumb camera dumping all data. This enables real-time, onboard decision-making for applications like disaster response, where detecting wildfires, floods, or structural damage minutes faster can save lives and property, all while operating within severe bandwidth constraints.