Satellite Embedding: A Review

Longhao Wang 

Satellite embeddings have become a practical interface between large-scale
Earth observation data and downstream geospatial analysis, yet the liter
ature is still organized mainly around foundation models rather than the
embeddings they produce. This review reframes the area from an embedding
centered perspective. We first define satellite embeddings as reusable latent
representations derived from satellite or multimodal Earth observation data,
and organize existing methods along three axes: model family, data and
modality support, and deployment mode. We then consolidate 19 represen
tative models into a unified dossier covering publication source, authorship,
spatial scale, temporal semantics, architecture, pretraining signal, embedding
dimensionality, runtime behavior, and code or artifact availability. Building
on the organization of recent remote sensing reviews, we connect the model
taxonomy to application requirements, data regimes, output granularity, and
operational constraints. We further report a compact Wuhan benchmark that
evaluates frozen embeddings on classification, fraction regression, water ex
traction, and dense segmentation within a shared 500m grid. The resulting
review is intended to serve as both a technical reference and an evidence
backed evaluation framework for selecting satellite embeddings in practical
Earth observation workflows.