Collaborative representation and confidence-driven semi-supervised learning for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is challenging due to spectral-spatial complexity and class imbalance.
• Existing methods often lack generalizability across diverse scenarios.

Purpose of the Study:

• To present a novel Graph-Convolutional Networks with Adaptive Region Ensembles (GCN-ARE) framework for robust HSI classification.
• To improve generalizability and address spectral-spatial complexity and class imbalance in HSI data.

Main Methods:

• Integrated graph spectral learning with a normalized graph Laplacian operator for stable feature propagation.
• Employed recursive K-means clustering under empirical risk bounds for adaptive region partitioning.
• Utilized theoretical guarantees (Hoeffding's inequality) for dynamic ensemble consistency and optimal classifier selection.

Main Results:

• GCN-ARE demonstrated superior performance over benchmarks like ViT and GAT on four HSI datasets.
• Achieved average Overall Accuracy (OA) improvements ranging from 1.5% to 5.7%.
• Ablation studies and parameter sensitivity analyses confirmed the efficacy and robustness of the adaptive subdivision and ensemble modules.

Conclusions:

• The GCN-ARE framework offers a theoretically rigorous and practically effective solution for robust HSI classification.
• The proposed methods enhance discriminability and consistency under spatial-spectral uncertainty.
• Sets a new standard for HSI classification performance and generalizability.