General-purpose compressed embeddings for geospatial foundation models

Develop methods to generate general-purpose compressed feature embeddings from geospatial foundation models that are sufficiently compact for storage and transmission yet retain utility across multiple downstream tasks (e.g., classification, segmentation, detection) without prior knowledge of task labels, by formulating appropriate rate–distortion objectives and training strategies for lossy neural compression of Earth Observation data.

Background

The paper reinterprets distortion from an algorithmic perspective, proposing feature compression where compressed representations are consumed directly by downstream models rather than reconstructed for human perception. It highlights that foundation models for vision often produce embeddings larger than the original data, making them impractical to store or transmit, and motivates compressing these features.

In this context, the authors explicitly identify the challenge of creating compressible, general-purpose embeddings that remain useful across tasks without a priori task supervision. Solving this would enable widespread distribution of ready-to-use features and democratize access to powerful models for Earth Observation analytics.