Trust-SSL: Additive-Residual Selective Invariance for Robust Aerial Self-Supervised Learning

Trust-SSL, developed by Wadii Boulila, Adel Ammar, Bilel Benjdira, and Maha Driss, tackles a key weakness in existing self-supervised learning (SSL) for aerial imagery: standard methods enforce invariance between augmented views even when one view is severely degraded by real-world corruptions like haze, motion blur, rain, or occlusion. This can introduce spurious structures into the latent space. Trust-SSL introduces a per-sample, per-factor trust weight into the alignment objective, combined with the base contrastive loss as an additive residual, rather than using a multiplicative gate that experiments showed impairs the backbone. The method uses a stop-gradient on the trust weight to preserve backbone quality.

Trained on a 210,000-image corpus over 200 epochs, Trust-SSL achieved the highest mean linear-probe accuracy among six backbones on EuroSAT, AID, and NWPU-RESISC45 (90.20%, vs. 88.46% for SimCLR and 89.82% for VICReg). Under severe information-erasing corruptions on EuroSAT, it gained +19.9 points on haze at severity 5 over SimCLR. The method also showed consistent +1 to +3 point gains in Mahalanobis AUROC on a zero-shot cross-domain stress test using BDD100K weather splits. Two ablations (scalar uncertainty and cosine gate) confirmed the additive-residual formulation as the primary improvement source. An evidential variant using Dempster-Shafer fusion adds interpretable signals of conflict and ignorance. Code is publicly available.