FA-Mamba: frequency attention driven Mamba for multimodal remote sensing classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Multimodal remote sensing classification is crucial for resource exploration and disaster monitoring.
• Existing Transformer and Mamba models face challenges with computational cost, long-range dependencies, and limited frequency-domain analysis.
• Noise and redundancy in multimodal data hinder effective feature fusion.

Purpose of the Study:

• To propose FA-Mamba, a frequency attention driven Mamba network for efficient and robust multimodal remote sensing classification.
• To enhance feature fusion by adaptively leveraging complementary information and suppressing noise across modalities.
• To improve the modeling of global contextual information and long-range correlations.

Main Methods:

• Developed FA-Mamba, a network with linear complexity for computational efficiency.
• Incorporated frequency attention to focus on relevant cross-modal information.
• Designed a coordinate fusion attention (CFA) mechanism to capture global spatial-temporal context and long-range dependencies.

Main Results:

• FA-Mamba demonstrates superior performance compared to state-of-the-art methods on three public datasets.
• Achieved an average overall accuracy of 95.84%.
• Effectively suppressed noise and redundancy, leading to more robust feature fusion.

Conclusions:

• FA-Mamba offers an efficient and effective solution for multimodal remote sensing classification.
• The frequency attention and CFA mechanisms significantly improve feature representation and classification accuracy.
• The approach provides a robust method for analyzing complex multimodal remote sensing data.