Few-shot remote sensing image scene classification based on multiscale covariance metric network (MCMNet).

Area of Science:

• Computer Science
• Remote Sensing
• Machine Learning

Background:

• Few-shot learning (FSL) enables rapid learning from limited data, crucial for remote sensing (RS) where annotated samples are scarce.
• Existing FSL methods often struggle with measuring inter-class similarity and intra-class variance in RS images due to Euclidean space embeddings.

Purpose of the Study:

• To propose a novel multi-scale covariance network (MCMNet) for enhanced remote sensing scene classification (RSSC).
• To address the limitations of Euclidean distance-based similarity measures in FSL for RS data.

Main Methods:

• Developed MCMNet using Conv64F backbone, mapping features from multiple layers (1, 2, 4) to manifold space via regional covariance matrices.
• Introduced manifold space prototypes for each feature layer to represent category centers.
• Implemented a multi-scale similarity measurement and three loss functions, optimized via an episodic training strategy.

Main Results:

• Comparative experiments on three public datasets demonstrated MCMNet's effectiveness.
• Achieved classification accuracy (CA) improvements ranging from 1.35% to 2.36% over state-of-the-art methods.
• The results validate MCMNet's superior performance in RSSC.

Conclusions:

• MCMNet effectively captures feature differences and similarities in manifold space for FSL in remote sensing.
• The proposed method offers a significant advancement for remote sensing scene classification with limited annotated data.