Naive Gabor Networks for Hyperspectral Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Convolutional Neural Networks (CNNs) excel in hyperspectral image (HSI) classification due to their representational power.
• High-dimensional HSI data and CNNs' large parameter counts necessitate substantial training samples to prevent overfitting.
• HSI classification with CNNs faces challenges like non-convex optimization, local minima, and flat regions.

Purpose of the Study:

• Introduce Gabor-Nets, a novel CNN architecture utilizing Gabor filters for HSI classification.
• Reduce the number of parameters and constrain the solution space in CNNs for improved HSI classification.
• Enhance the performance of CNNs in HSI classification, particularly when training data is scarce.

Main Methods:

• Design and learn CNN kernels strictly in the form of Gabor filters, termed Gabor-Nets.
• Develop an innovative phase-induced Gabor kernel for feature learning through data component combinations.
• Implement complex-valued Gabor filtering in a real-valued manner for seamless integration into standard CNN frameworks.

Main Results:

• Gabor-Nets significantly improve CNN performance in HSI classification.
• The proposed method demonstrates superior results, especially when using a small training set.
• Gabor-Nets effectively adapt to local harmonic characteristics of HSI data, yielding representative harmonic features.

Conclusions:

• Gabor-Nets offer a parameter-efficient and effective approach for HSI classification.
• The phase-induced Gabor kernel enables adaptive feature learning and robust performance.
• This Gabor filter-based CNN design addresses key challenges in HSI analysis, particularly with limited data availability.