CESA-MCFormer: An Efficient Transformer Network for Hyperspectral Image Classification by Eliminating Redundant...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification faces challenges due to spectral and spatial redundancy.
• Applications in agriculture and infrastructure detection require robust feature extraction.

Purpose of the Study:

• To propose an innovative model, CESA-MCFormer, for effective HSI classification.
• To address information redundancy in HSI data for improved classification accuracy.

Main Methods:

• Developed the CESA-MCFormer model based on the transformer architecture.
• Introduced the Center Enhanced Spatial Attention (CESA) module for prior spatial information.
• Incorporated Morphological Convolution (MC) for feature extraction and merging.

Main Results:

• CESA-MCFormer achieved high classification performance on IP, UP, and Chikusei datasets.
• Kappa coefficients reached 96.38%, 98.24%, and 99.53% respectively.
• The model demonstrated precise classification with minimal samples, without PCA.

Conclusions:

• The CESA-MCFormer model offers a powerful solution for HSI classification.
• The integration of CESA and MC modules enhances feature processing capabilities.
• This method provides a significant advancement in HSI analysis for various applications.