Mixed convolutional classification method for hyperspectral images based on spatial spectrum orthogonal constraints...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• 3D-CNN and 2D-CNN models struggle with hyperspectral image classification due to high computational costs and feature extraction challenges.
• Mixed convolution offers improvements but suffers from information loss/redundancy and reduced accuracy on small datasets.
• Current feature compression methods like max pooling overlook feature interdependencies.

Purpose of the Study:

• To develop an advanced hyperspectral image classification method that addresses the limitations of existing CNN architectures.
• To enhance feature extraction and reduce computational resource consumption in hyperspectral image analysis.
• To improve classification accuracy, particularly in scenarios with limited sample data.

Main Methods:

• A novel parallel convolution architecture is proposed.
• Spatial spectral orthogonal constraints and a bidirectional attention mechanism are introduced to ensure information independence and enhance key feature utilization.
• A combination of max pooling and covariance pooling is used for feature compression, balancing salient and high-order statistical features.
• The Mish activation function replaces ReLU for improved stability in data output.

Main Results:

• The proposed method achieved high Overall Accuracy (OA) values of 93.3%, 94.1%, and 99.1% on three public datasets.
• Demonstrated significant superiority over commonly used hyperspectral image classification methods.
• The approach effectively extracts discriminative features, leading to improved classification performance.

Conclusions:

• The novel parallel convolution architecture offers a powerful solution for hyperspectral image classification.
• The integration of spatial spectral orthogonal constraints, bidirectional attention, and advanced pooling strategies enhances feature representation and classification accuracy.
• This method provides a robust and efficient alternative for hyperspectral image analysis, especially in challenging conditions like small sample sizes.