Adapting self-supervised pretraining to radio astronomy imaging constraints

Determine how to adapt self-supervised pretraining methods—specifically masked image modeling and contrastive or invariance-based representation learning—to radio astronomy imaging characterized by single-channel inputs, survey-dependent intensity distributions, and heterogeneous instrumental and imaging systematics, so that the learned representations transfer effectively across telescopes and imaging pipelines.

Background

The paper situates STRADAViT within the growing use of self-supervised pretraining in computer vision, where methods such as masked image modeling and contrastive objectives have produced strong transferable features, particularly with Vision Transformers. However, radio astronomy data differ substantially from natural images: they are typically single-channel, exhibit survey-specific intensity distributions, and contain heterogeneous instrumental and imaging artefacts.

Motivated by these domain constraints, the authors propose STRADAViT with mixed-survey pretraining data, radio-aware view generation, and staged reconstruction-to-contrastive objectives. The explicit open question highlights the need for principled adaptation of these self-supervised techniques to the scientific-imaging regime of radio astronomy, beyond ad hoc modifications, to ensure robust cross-survey transfer.