GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation

GeoSET. Generalist Foundation Model for SAR-to-EO Image Translation. By Jeonghyeok Do and one other, from KAIST. Published on the twenty-sixth of September. Its code repository is worked out: about eighty source files, with a readme and a licence.

Synthetic aperture radar and electro-optical imagery now come in pairs across many sensors, many resolutions and many geographic regions. The data has spread out. The models have not. Existing methods for translating radar images into electro-optical ones, a task called SET, are typically trained on a single dataset of limited scale, and what comes out is a model specialised to the particular sensing conditions of that dataset.

GeoSET is offered as the first generalist model for this task. It is built around a single pretrained parent, and that parent is then adapted to each downstream dataset under one common protocol. Specialisation still happens, but it happens after a broad pretraining rather than in place of one.

The pretraining corpus is curated at scale. From a collection of more than ten million SAR observations, the authors keep over three million high-quality SAR and EO pairs, spanning diverse sensors, spatial resolutions and ground sampling distances. Radar observations and a pretrained image generator sit on opposite sides of a modality gap, so to bridge it they develop a speckle-robust SAR encoder, and they pretrain the conditional generator on this heterogeneous corpus.

Adaptation is meant to be cheap. The parent is adapted with low-rank adaptation, LoRA, which updates only zero point six percent of the generator's parameters and takes approximately one hour per dataset.

Six downstream benchmarks are used. On them, GeoSET achieves state-of-the-art results on two measures, FID and DISTS, and it does so whether the parent is fully fine-tuned or adapted with LoRA. The authors read this as effective transfer across heterogeneous SAR and EO domains.

What carries the result is the order of operations. Breadth comes first, in one pretraining corpus that spans sensors and resolutions. Specialisation comes second, at the cost of a sliver of the parameters and about an hour for each dataset. On six benchmarks, that order reaches the state of the art.