A Spectral-Spatial-Dependent Global Learning Framework for Insufficient and Imbalanced Hyperspectral Image...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning excels in hyperspectral image (HSI) classification but requires substantial labeled data.
• Existing patch-based and patch-free methods struggle with imbalanced HSI datasets.
• Fast patch-free global learning (FPGA) architecture faces challenges in extracting discriminative features from imbalanced samples.

Purpose of the Study:

• To propose a spectral-spatial-dependent global learning (SSDGL) framework for HSI classification with insufficient and imbalanced data.
• To address the limitations of existing deep learning models in handling imbalanced HSI datasets.
• To enhance feature extraction capabilities for improved HSI classification accuracy.

Main Methods:

• Developed a spectral-spatial-dependent global learning (SSDGL) framework incorporating global convolutional long short-term memory (GCL) and global joint attention mechanism (GJAM).
• Introduced a hierarchically balanced (H-B) sampling strategy and weighted softmax loss to mitigate data imbalance issues.
• Utilized GCL for long short-term spectral feature dependency extraction and GJAM for learning discriminative feature representations.

Main Results:

• The proposed SSDGL framework demonstrated superior performance on three public HSI datasets.
• SSDGL effectively handled insufficient and imbalanced sample problems in HSI classification.
• Experimental results show significant improvements over existing state-of-the-art methods.

Conclusions:

• The SSDGL framework offers a powerful solution for HSI classification, particularly in scenarios with limited and imbalanced data.
• The integration of GCL and GJAM modules enhances the model's ability to extract discriminative spectral-spatial features.
• The proposed methods provide a robust approach for improving HSI classification accuracy and overcoming data challenges.