Transferability of self-supervised hyperspectral models

Determine whether self-supervised masked-autoencoder models transfer effectively to small-sample downstream hyperspectral image classification without overfitting.

Background

The paper reviews self-supervised masked-autoencoder approaches for hyperspectral imaging, including methods that reconstruct masked spectral-spatial tokens, operate in the frequency domain, or perform foundation-scale pretraining across multiple data sources. Although such pretraining is motivated by the expense of hyperspectral annotation, the paper identifies an unresolved issue concerning downstream transfer. Specifically, it is not established whether representations learned without labels remain useful for classification when only a small number of labelled samples are available, or whether the models instead overfit during fine-tuning.

References

Whether such models transfer to small-sample downstream classification without overfitting is open, and it is a question a controlled harness is well placed to examine.

— Hyperspectral Image Models: Technical Report  (2609.39871 - Rachamalla et al., 30 Sep 2026) in Section 2.1, paragraph “Self-supervised pretraining”

While the mean gain was in AEF's favor for both settings, we consider the extreme degree of variability indicative that adequate general 1-shot or 10-shot performance remains an unsolved research frontier.

— AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data  (2507.22291 - Brown et al., 29 Jul 2025) in Section S6.1, “Comparisons with 10 and 1 samples per-class”