Small Sample Hyperspectral Image Classification Based on the Random Patches Network and Recursive Filtering.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning frameworks for hyperspectral image (HSI) classification often exhibit high complexity and low accuracy in few-shot learning scenarios.
• Existing methods struggle to achieve optimal performance when training data is scarce.

Purpose of the Study:

• To develop an effective HSI classification method that overcomes the limitations of complex models and few-shot learning.
• To enhance classification accuracy by extracting informative deep features from HSI data.

Main Methods:

• A novel method combining random patches network (RPNet) for multi-level feature extraction and recursive filtering (RF) for feature refinement.
• Utilized principal component analysis (PCA) for dimension reduction of RPNet features.
• Integrated spectral features with RPNet-RF features for classification using a support vector machine (SVM).

Main Results:

• The proposed RPNet-RF method demonstrated superior performance on three benchmark HSI datasets with limited training samples.
• Achieved higher overall accuracy and Kappa coefficient values compared to advanced HSI classification methods designed for small training sets.
• Effectively extracts informative deep features crucial for accurate HSI classification.

Conclusions:

• The RPNet-RF method offers a robust and accurate solution for hyperspectral image classification, particularly in few-shot learning contexts.
• This approach provides a significant advancement in handling complex HSI data with limited labeled samples.
• The combination of RPNet and RF effectively enhances feature representation for improved classification outcomes.