Cross-Image Federated Learning for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Earth Observation
• Machine Learning

Background:

• Modern remote sensing increasingly uses multisatellite and multiplatform Earth observation data.
• Traditional single-image processing (SIP) for hyperspectral images (HSIs) limits model generalization across spatial and temporal domains.
• Growing complexity in HSI applications highlights SIP's limitations.

Purpose of the Study:

• Propose a cross-image hyperspectral image federated learning approach for classification.
• Enhance personalization and learning efficiency for individual clients.
• Address global knowledge bias from uneven data distribution in federated learning.

Main Methods:

• Developed a client-oriented self-guided knowledge-enhanced personalized learning method.
• Introduced a multiscale semantic aligned dynamic aggregation method for fair global knowledge integration.
• Constructed open-set and closed-set datasets for HSI classification using federated learning.

Main Results:

• Demonstrated the effectiveness of the proposed federated learning approach on tailored datasets.
• Improved learning efficiency and personalization for clients by leveraging inter-client features.
• Ensured fairness in global knowledge aggregation despite uneven data distribution.

Conclusions:

• This work pioneers federated learning for joint hyperspectral image classification.
• The proposed methods effectively enhance HSI classification generalization and efficiency.
• The approach addresses key challenges in distributed HSI data analysis.