ATLAS: Evidence-Gated Annotation Transport for Satellite Semantic Segmentation

Satellite semantic segmentation must accommodate appearance variations across geography, urbanization, season, and sensor. A fixed parametric classifier can compromise across these modes, while retrieval-based prediction depends on the availability of relevant labeled support. We propose ATLAS, an evidence-gated annotation-transport framework that combines a parametric segmentation head with a complementary nonparametric expert. The transport expert retrieves ground-truth-labeled, mode-indexed region atoms from the training corpus, conditioning its predictions on the appearance mode of the current scene rather than relying on self-generated pseudo-labels. A per-pixel gate with two learned scalars uses retrieval affinity to regulate the contribution of transported annotations. The deployed posterior is $P(i)=(1-w(i))P^s(i)+w(i)P^E(i)$, where $P^s$ and $P^E$ denote the parametric and transport posteriors. This construction bounds the deviation from the parametric prediction by $|P(i)-P^s(i){|}_{\mathrm{TV}}\le w(i)$ and recovers the parametric head exactly when retrieval returns no support. Thus, the gate controls the influence of retrieval without requiring it to be reliable everywhere. During training, the transport expert also supervises the parametric head through distillation, aiming to internalize mode-conditional information in the fallback. Setting $w\equiv 0$ removes the atlas, retrieval, and projector, enabling memory-free deployment with the baseline architecture and no additional inference overhead. Conversely, retaining the atlas allows new-mode labeled regions to be incorporated without retraining. We outline an evaluation on LoveDA and ISPRS Potsdam/Vaihingen targeting state-of-the-art accuracy through matched-backbone comparisons, compute-matched training, and ensemble controls, with both gated and memory-free operating points reported to distinguish the value of annotation transport from additional model capacity or computation.