Author(s): Amrith Coumaran

Originally published on Towards AI.

Zero-Shot SAR-Optical Matching

Can pretrained vision models match satellite images from different sensors without any training?

The article investigates zero-shot matching between Synthetic Aperture Radar (SAR) and optical satellite imagery using five pretrained feature extractors (SD-VAE, SD-UNet, DINOv2, CLIP, and SAM2). It introduces experiments measuring (1) modality gap via cosine distance, showing SAM2 achieves the strongest alignment, (2) patch retrieval accuracy, where retrieval performance does not follow alignment—SD-VAE performs best despite having the worst modality gap, and (3) feature-space structure using UMAP, heatmaps, clustering, and separability analysis to reveal the core tension: strong modality invariance (SAM2) can come at the cost of discriminative power needed to distinguish specific patches, while spatially faithful representations (SD-VAE) enable better matching. The discussion highlights why absolute recall is low under the hard, zero-shot, single-polarization, co-registered, patch-level setting, and concludes with practical guidance: use a two-stage approach combining SAM2 for coarse candidate narrowing and SD-VAE for fine-grained verification.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI