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

Research Question and Approach

Multi-label remote sensing (RS) classification faces significant challenges when models are deployed on data distributions that differ from their training sets. While existing style-augmentation methods (like MixStyle or EFDMix) improve generalization by perturbing feature statistics, they typically treat an entire image as a single style. In multi-label aerial scenes—where diverse objects like water, buildings, and vegetation coexist—this global approach causes class contamination, where the style of one object is incorrectly applied to another.

To solve this, the authors propose a label-decoupled (LD) augmentation framework. This method uses spatial attention to isolate the feature statistics of individual labels. These per-label statistics are then mixed with cross-domain samples that share the same labels, using independent coefficients for each class. The features are then recomposed using attention-weighted normalization, ensuring that style perturbations are confined to the relevant semantic regions.

Key Findings

The researchers evaluated six variants of their framework across a new leave-one-domain-out benchmark constructed from the UCM, AID, and DFC15 datasets. The best-performing variant achieved a mean average precision of 71.5%, representing a 5.0-point improvement over standard empirical risk minimization and a 1.3-point gain over the strongest global-statistics baseline. Notably, the framework provided the most significant improvements on the hardest transfer tasks, with gains reaching up to 7.7 points. Ablation studies confirmed that spatial attention and the use of refreshed localization maps are the primary drivers of this performance.

This study provides a practical, low-cost upgrade for multi-label remote sensing models. The framework adds at most 0.35% to the total parameter count and requires no changes to the model during inference, making it highly efficient for real-world deployment. By addressing the specific structural nuances of multi-label aerial imagery, the authors offer a robust solution for domain generalization that avoids the pitfalls of generic, image-level augmentation techniques.