Seed-to-Semantics: Few-Shot Prototype-Guided Progressive Learning for Hyperspectral and LiDAR Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning fusion of hyperspectral images (HSI) and LiDAR excels in remote sensing classification.
• High costs of pixel-wise annotation limit its application in label-scarce settings.
• Conventional deep models overfit, and standard semi-supervised learning (SSL) methods suffer from confirmation bias in few-shot scenarios.

Purpose of the Study:

• To develop a robust framework for few-shot HSI-LiDAR classification in extremely label-scarce regimes.
• To overcome overfitting and confirmation bias issues in existing deep and semi-supervised learning methods.
• To improve the reliability of pseudo-labels during both initialization and self-training.

Main Methods:

• Propose Prototype-Guided Progressive Learning (PGPL), a unified framework for few-shot HSI-LiDAR classification.
• Construct a reliable initialization pool using spectral-angle and elevation-consistency cues in the original data domain.
• Progressively expand the training set via class-balanced pseudo-label admission and temporal confidence stabilization.

Main Results:

• PGPL consistently outperforms state-of-the-art supervised and semi-supervised baselines in 2-5-shot settings.
• Achieved accuracy gains of 4.64% (Houston), 1.16% (Trento), and 3.92% (MUUFL).
• Demonstrated higher pseudo-label purity compared to competing methods.

Conclusions:

• PGPL offers a superior solution for few-shot HSI-LiDAR classification, especially in data-limited environments.
• The framework effectively addresses challenges of overfitting and confirmation bias.
• PGPL significantly advances the performance of multimodal remote sensing classification under extreme label scarcity.