S-MAT: Semantic-Driven Masked Attention Transformer for Multi-Label Aerial Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Multi-label aerial scene image classification is challenging due to co-existing land cover objects.
• Existing methods struggle with modeling dependencies involving non-existent categories, leading to performance degradation.

Purpose of the Study:

• To propose S-MAT, a Semantic-driven Masked Attention Transformer, for robust multi-label aerial scene image classification.
• To enhance the modeling of label dependencies by filtering redundant information.

Main Methods:

• S-MAT utilizes a Masked Attention Transformer (MAT) to capture correlations among label embeddings.
• A Semantic Disentanglement Module (SDM) constructs label embeddings.
• Masked attention mechanism filters redundant dependencies, improving model robustness.

Main Results:

• Achieved CF1 scores of 89.21% (UC-Merced), 90.90% (AID), and 88.31% (MLRSNet).
• Demonstrated effectiveness through extensive ablation studies and empirical analysis.

Conclusions:

• S-MAT effectively captures label dependencies, outperforming previous methods.
• The proposed masked attention mechanism enhances classification robustness and accuracy in aerial imagery.