Label-Decoupled Style Augmentation for Domain Generalization in Multi-Label Remote Sensing Scene Classification

2026-07-14 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Remote Sensing · Depth: Expert, quick

Summary

A novel Label-Decoupled Style Augmentation framework is proposed to enhance domain generalization in multi-label remote sensing scene classification, addressing the issue where global feature-statistics augmentation contaminates classes. This framework confines style perturbation to label-specific regions by using per-label attention, derived from a learnable module or gradient class-activation maps. Per-label feature statistics are then mixed with cross-domain samples sharing present labels, using independent per-label coefficients, and features are recomposed via attention-weighted normalization. Evaluated on a leave-one-domain-out benchmark across multi-label UCM, AID, and DFC15 datasets with six shared labels, the best variant achieved 71.5% mean average precision. This result surpasses empirical risk minimization by 5.0 points and the strongest global-statistics baseline by 1.3 points, with gains up to 7.7 points on the hardest transfers. The framework adds at most 0.35% parameters and leaves inference unchanged.

Key takeaway

For Machine Learning Engineers developing robust multi-label remote sensing classification models, you should consider integrating label-decoupled style augmentation. This approach significantly boosts domain generalization, particularly for challenging transfers, achieving up to 7.7 points gain on hard transfers. It offers an inexpensive upgrade path, adding at most 0.35% parameters and leaving inference unchanged, making it a practical solution for improving model resilience to distribution shifts.

Key insights

Label-decoupled style augmentation improves multi-label domain generalization by localizing style perturbations to specific class regions.

Principles

• Global style augmentation can lead to class contamination.
• Confining style perturbation to label-specific regions is effective.
• Spatial attention is key for localized feature mixing.

Method

The framework uses per-label attention to obtain label-specific feature statistics, mixes these with cross-domain samples sharing labels under independent coefficients, and recomposes features via attention-weighted normalization.

In practice

• Implement per-label attention for fine-grained style mixing.
• Integrate this framework for multi-label remote sensing tasks.
• Utilize gradient class-activation maps for per-label attention.

Topics

• Label-Decoupled Augmentation
• Domain Generalization
• Multi-Label Classification
• Remote Sensing
• Feature Augmentation
• Attention Mechanisms
• Mean Average Precision

Code references

• Alaa-Almouradi/Style-Augmentation-Upgrade

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• YNU-HPCC at SemEval-2026 Task 9: Hybrid Augmentation and Regularization Strategies for Multilingual Polarization Type Classification — A hybrid augmentation and regularization framework improves multilingual polarization classification on imbalanced data using XLM-RoBERTa-large.
• Domain-Incremental Remote Sensing Change Detection via Difference-Guided Adaptation and Frequency-Decoupled Distillation — DG-FDD mitigates catastrophic forgetting in remote sensing change detection by integrating difference-guided adaptation and frequency-decoupled distillation.
• Anti-causal domain generalization: Leveraging unlabeled data — In anti-causal settings, utilizing unlabeled data from multiple environments can improve domain generalization by penalizing model sensitivity to covariate shifts.
• Complex Layout Classification in the Wild: A Low-Resource Approach with Layout-Preserving Augmentations — Domain-aware augmentations significantly improve low-resource complex layout classification by focusing on global geometric arrangements.
• Semantically Aware Optimal Transport for Dense Label Transfer — SAOT improves dense label transfer across visual domains by learning semantic transport costs with unbalanced optimal transport.
• Adversarial Label Invariant Graph Data Augmentations for Out-of-Distribution Generalization — RIA improves OoD generalization under covariate shift using adversarial label-invariant data augmentations.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.