CSA-Kansformer : Cross-scale aggregation and Kansformer network for hyperspectral image classification

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Hyperspectral image (HSI) classification is crucial for remote sensing, requiring models that can effectively utilize rich spectral and spatial information.
• Convolutional neural networks (CNNs) excel at local feature extraction, while transformers capture global context, but each has limitations for HSI data.
• Existing methods often struggle to balance feature extraction capabilities with computational efficiency for complex HSI datasets.

Purpose of the Study:

• To introduce a novel hybrid model, CSA-Kansformer, that combines the strengths of CNNs and transformers for improved HSI classification.
• To enhance feature representation and reduce computational load through innovative modules like SCConv, CSAM, and an optimized Kansformer block.
• To achieve state-of-the-art performance in HSI classification tasks with greater efficiency.

Main Methods:

• Developed the CSA-Kansformer model featuring a spatial and channel reconstruction convolution (SCConv) block for feature reduction and abstract spatial-spectral feature extraction.
• Incorporated a cross-scale aggregation module (CSAM) with fusion convolution, channel attention, and spatial attention for efficient multi-scale feature aggregation.
• Introduced an optimized Kansformer block utilizing batch normalization and Kolmogorov-Arnold Networks (KANs) for improved training stability, convergence, and performance.

Main Results:

• Extensive experiments on four benchmark HSI datasets (Botswana, Houston2013, WHU-Hi-HanChuan, WHU-Hi-HongHu) demonstrated superior performance of the CSA-Kansformer model.
• The proposed model significantly outperformed nine state-of-the-art methods in terms of classification accuracy.
• CSA-Kansformer achieved notable improvements in computational efficiency compared to existing approaches.

Conclusions:

• The CSA-Kansformer model effectively addresses the limitations of traditional CNNs and transformers in HSI classification.
• The integration of SCConv, CSAM, and the optimized Kansformer block leads to enhanced feature representation, accuracy, and efficiency.
• This research offers a promising direction for developing advanced deep learning models for hyperspectral image analysis in remote sensing.