# Cross-modal learning for SAR target recognition using optical vision foundation models

**Authors:** Lucas Hirsch, James R. Hopgood, Javid Khan, Yoann Altmann, Mike E. Davies  
**arXiv:** [2609.07753](https://arxiv.org/abs/2609.07753)  
**Date:** 2026-09-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2609.07753v1/

## Key points

- Synthetic Aperture Radar (SAR) is known for its versatile, long-range capabilities and ability to operate in almost all weather conditions.
- Automatic Target Recognition (ATR) in SAR images faces challenges due to limited labeled data, strong speckle interference, and a significant domain gap with optical imagery.
- Electro-optical (EO) imagery benefits from extensive datasets, clearer visual structure, and robust foundation models.
- The study explores using vision foundation models trained on optical data to provide class-level supervision for SAR classification.
- A cross-modal EO to SAR prototype alignment framework is proposed where a frozen EO encoder constructs class-level optical prototypes without strict EO/SAR pairs.
- The approach enhances SAR classification accuracy compared to other baselines and offers clearer class separation within the trained SAR embedding space.

## Summary

, , , , 

Synthetic Aperture Radar (SAR) is a crucial imaging modality known for its versatile, long-range capabilities and ability to operate in almost all weather conditions. However, Automatic Target Recognition (ATR) in SAR images poses challenges due to limited labeled data, strong speckle interference, and a significant domain gap between SAR and more abundant optical imagery. In contrast, electro-optical (EO) imagery benefits from extensive datasets, clearer visual structure, and robust foundation models. This study explores the potential of using vision foundation models trained on optical data to provide class-level supervision for SAR classification. The researchers propose a cross-modal EO to SAR prototype alignment framework where a frozen EO encoder, based on a DINOv3 vision foundation model, constructs class-level optical prototypes without strict EO/SAR pairs. Subsequently, a SAR model is trained to classify SAR images while aligning its embeddings with the corresponding EO class prototype. Importantly, at inference time, the SAR model operates independently without access to optical imagery. The effectiveness of this approach is evaluated on the UNICORNv2 dataset—a collection of civilian vehicle EO and SAR images characterized by heavy speckling and severe class imbalance. The results demonstrate that EO prototype alignment enhances SAR classification accuracy compared to frozen DINOv3, SAR-only fine-tuning, and unpaired distribution alignment baselines. Additionally, qualitative evidence from t-SNE visualizations indicates clearer class separation within the trained SAR embedding space. Overall,<kgd> this study suggests that optical vision foundation models—despite being trained on visible spectrum imagery—offer transferable information for improving SAR image classification.</kgd> This novel approach presents a practical method for leveraging large-scale pretrained vision foundation models across challenging sensing modalities like SAR imaging.