Contextual Learning in Fourier Complex Field for VHR Remote Sensing Images

Area of Science:

• Remote Sensing (RS)
• Computer Vision
• Artificial Intelligence

Background:

• Very high-resolution (VHR) remote sensing image classification is crucial for image analysis.
• Transformer models excel at capturing contextual relationships but are computationally intensive for VHR images.
• Quadratic complexity of standard Transformers limits their application to large-scale VHR remote sensing data.

Purpose of the Study:

• To develop an efficient Transformer-based model for VHR remote sensing image classification.
• To address the computational challenges posed by large image sizes in VHR remote sensing.
• To enhance the learning of high-order contextual information in remote sensing images.

Main Methods:

• Proposed an efficient complex self-attention (CSA) mechanism by decomposing self-attention using discrete Fourier transform (DFT).
• Utilized the conjugated symmetric property of DFT for reduced computational complexity (less than half of naive SA).
• Introduced a Logmax function to stabilize gradient propagation and normalize attention maps in the complex field, forming the Fourier Complex Transformer (FCT).

Main Results:

• The Fourier Complex Transformer (FCT) model effectively learns global contextual information hierarchically from VHR aerial images.
• Experiments demonstrate the effectiveness and efficiency of FCT, particularly for VHR remote sensing image classification tasks.
• FCT achieves comparable or superior performance to existing methods with significantly reduced computational cost.

Conclusions:

• The proposed CSA mechanism and FCT model offer an efficient solution for VHR remote sensing image classification.
• FCT successfully overcomes the computational limitations of standard Transformers for high-resolution imagery.
• The approach provides a promising direction for advancing remote sensing image analysis and understanding.