Optimal serialization for state-space models

Determine which one-dimensional serialization of two-dimensional hyperspectral image neighborhoods most effectively supports state-space models, particularly with respect to classification accuracy and scan-induced artifacts.

Background

Selective state-space models process sequences and therefore require a one-dimensional ordering when applied to two-dimensional hyperspectral patches. The paper observes that different state-space architectures use different scan or serialization strategies, and that the family exhibits substantial variation in performance and visible directional streaking. Because the 11×1111\times11 patches used in the benchmark contain only 121 tokens, the asymptotic efficiency advantage of linear-time scanning is limited, making the choice of ordering itself a central unresolved issue. The paper therefore leaves open which serialization design best preserves spatial information and improves model behavior.

References

The state-space family reaches the accuracy of the best transformers at its top end while being the most dispersed family overall, and the streaking in \Cref{fig:maps} points at why. A selective scan needs a one-dimensional ordering of a two-dimensional neighbourhood, and at an $11\times11$ window the sequence is only 121 tokens long, so the asymptotic advantage of linear-time scanning has little room to express itself while the cost of choosing an ordering remains. Variation within the family tracks how each design handles that choice more closely than it tracks publication date, which suggests scan design rather than scan efficiency is where the remaining gains are.

— Hyperspectral Image Models: Technical Report  (2609.39871 - Rachamalla et al., 30 Sep 2026) in Section 6.7, paragraph “Serialisation is the open question for state-space models”