Rotation-Invariant Attention Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification uses spectral and spatial information.
• Deep learning methods often use 3x3 convolutions, which are sensitive to spatial rotation.
• This sensitivity degrades performance on rotated HSIs.

Purpose of the Study:

• To propose a novel rotation-invariant attention network (RIAN) for improved HSI classification.
• To address the limitations of existing methods in handling rotated HSI data.
• To enhance the accuracy and robustness of land-cover identification in HSIs.

Main Methods:

• Developed a center spectral attention (CSpeA) module to suppress redundant spectral bands.
• Introduced a rectified spatial attention (RSpaA) module to replace 3x3 convolution for rotation-invariant feature extraction.
• Integrated CSpeA, 1x1 convolution, and RSpaA to construct the RIAN.

Main Results:

• The proposed RIAN demonstrates invariance to spatial rotation in HSIs.
• RIAN achieved superior performance, with an overall accuracy of 86.53% on the Houston database, a 1.04% improvement.
• Experimental results validate the effectiveness of RIAN in HSI classification.

Conclusions:

• RIAN effectively overcomes the challenge of spatial rotation in HSI classification.
• The network offers a robust solution for accurate land-cover identification.
• The developed method shows significant potential for real-world HSI applications.