AI · 8 min read · April 24, 2026
Trust-weighted SSL improves aerial image learning under corruption
Additive-residual trust weights boost self-supervised learning robustness when aerial images degrade, outperforming standard contrastive methods on benchmark datasets.
Trust-SSL adds per-sample confidence weights to contrastive loss, improving aerial image representation learning under haze, blur, and occlusion.
- — Standard SSL enforces alignment between clean and degraded views, risking spurious latent structure.
- — Trust-SSL assigns per-sample, per-corruption-factor weights to the alignment objective.
- — Additive-residual formulation with stop-gradient outperforms multiplicative gating on backbone quality.
- — Achieves 90.20% linear-probe accuracy on EuroSAT, AID, NWPU-RESISC45 benchmarks.
- — Gains +19.9 points on severe haze corruption versus SimCLR baseline.
- — Evidential variant using Dempster-Shafer fusion provides interpretable conflict and ignorance signals.
- — Cross-domain stress test shows +1 to +3 point improvements in Mahalanobis AUROC.
Frequently asked
- Multiplicative gating can suppress gradient flow through the backbone when trust weights are low, degrading feature learning. Additive-residual formulation preserves the base contrastive signal while adding a confidence-scaled correction term, allowing the backbone to learn from both clean and degraded samples without bottlenecking gradients. Experiments show this design yields higher linear-probe accuracy and robustness.