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.
Astrobobo tool mapping
- Knowledge Capture Document the additive-residual principle: uncertainty signals should be added to loss terms, not gated multiplicatively, to preserve backbone gradient flow.
- Focus Brief Summarize the three key ablations (scalar uncertainty, cosine gate, additive-residual) to understand which design choice drives robustness gains.
- Reading Queue Queue related papers on uncertainty-aware contrastive learning and Dempster-Shafer fusion for deeper context on evidential SSL variants.
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.
cite ▸
APA
Wadii Boulila, Adel Ammar, Bilel Benjdira, Maha Driss. (2026, April 24). Trust-weighted SSL improves aerial image learning under corruption. Astrobobo Content Engine (rewrite of arxiv/cs.AI). https://astrobobo-content-engine.vercel.app/article/trust-weighted-ssl-improves-aerial-image-learning-under-corruption-0f18dc
MLA
Wadii Boulila, Adel Ammar, Bilel Benjdira, Maha Driss. "Trust-weighted SSL improves aerial image learning under corruption." Astrobobo Content Engine, 24 Apr 2026, https://astrobobo-content-engine.vercel.app/article/trust-weighted-ssl-improves-aerial-image-learning-under-corruption-0f18dc. Based on "arxiv/cs.AI", https://arxiv.org/abs/2604.21349.
BibTeX
@misc{astrobobo_trust-weighted-ssl-improves-aerial-image-learning-under-corruption-0f18dc_2026,
author = {Wadii Boulila, Adel Ammar, Bilel Benjdira, Maha Driss},
title = {Trust-weighted SSL improves aerial image learning under corruption},
year = {2026},
url = {https://astrobobo-content-engine.vercel.app/article/trust-weighted-ssl-improves-aerial-image-learning-under-corruption-0f18dc},
note = {Astrobobo rewrite of arxiv/cs.AI, https://arxiv.org/abs/2604.21349},
}