ARTFEED — Contemporary Art Intelligence

Trust-SSL: Enhancing Robustness in Aerial Self-Supervised Learning

ai-technology · 2026-04-25

A recent study presents Trust-SSL, a novel training approach and architectural adjustment aimed at enhancing self-supervised learning (SSL) for aerial images. Traditional SSL techniques focus on maintaining consistency between augmented views; however, aerial imagery frequently faces challenges such as haze, motion blur, rain, and occlusion that obscure vital details. Aligning clean and degraded views can lead to misleading structures. Trust-SSL incorporates a trust weight for each sample and factor into the alignment goal, which is added to the base contrastive loss. Instead of using a multiplicative gate, a stop-gradient is applied to the trust weight, as experiments indicate that the latter weakens the backbone. This additive-residual method boosts robustness. The findings are available on arXiv as preprint 2604.21349.

Key facts

  • Trust-SSL is a new training strategy for aerial self-supervised learning.
  • It addresses corruptions like haze, motion blur, rain, and occlusion.
  • A per-sample, per-factor trust weight is introduced into the alignment objective.
  • The trust weight is combined with the base contrastive loss as an additive residual.
  • A stop-gradient is applied to the trust weight instead of a multiplicative gate.
  • Multiplicative gate impairs the backbone, while additive-residual improves robustness.
  • The study is published on arXiv as preprint 2604.21349.
  • The paper is categorized as cross-type announcement.

Entities

Institutions

  • arXiv

Sources