Adaptive pixel attention network for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Fixed convolution kernels in hyperspectral image (HSI) classification limit channel weight learning.
• This limitation hinders the capture of inter-pixel connections within patches, negatively impacting classification performance.

Purpose of the Study:

• To propose a novel Adaptive Pixel Attention Network for improved HSI classification.
• To enhance the mining of pixel connections within patch features for better performance.

Main Methods:

• Developed a Spectral-Spatial Superposition Enhancement module to enrich spectral-spatial information.
• Introduced an Adaptive Pixel Attention mechanism using Cosine and Euclidean similarity to analyze pixel relationships across different scales.
• Implemented a Cross-Layer Information Complement module for contextual interaction and to prevent information omission.

Main Results:

• The proposed network demonstrated superior accuracy compared to state-of-the-art models on four HSI datasets (IP, UP, HU, KSC).
• Achieved better efficiency than existing 3D hyperspectral image classification methods.

Conclusions:

• The Adaptive Pixel Attention Network effectively improves HSI classification by adaptively learning pixel relationships.
• The proposed modules enhance spectral-spatial information and contextual interactions, leading to superior performance and efficiency.