Optical-Guided Neural Collapse for SAR Few-Shot Class Incremental Learning

Researchers propose an optical-guided neural collapse framework for SAR few-shot class incremental learning that addresses data scarcity and catastrophic forgetting by transferring geometric structure from optical imagery to SAR domain. The method achieves superior performance on benchmark datasets while maintaining better feature compactness and inter-class separability compared to existing FSCIL approaches.

This research addresses a specialized challenge in synthetic aperture radar (SAR) image analysis where limited training data and sequential learning create significant accuracy degradation. The core innovation transfers discriminative structure from optical automated target recognition (ATR) datasets to SAR imagery by leveraging orthogonal feature subspaces as geometric priors. This cross-domain transfer approach tackles the inherent difficulty of SAR's azimuth sensitivity, which creates substantial intra-class variation and inter-class confusion—problems that existing few-shot learning methods struggle to resolve.

The breakthrough centers on neural collapse geometry, a phenomenon where neural network features naturally organize into mathematically optimal configurations. By constraining SAR features to project onto optical-derived subspaces through principal angle constraints, the researchers induce this collapse structure while preventing catastrophic forgetting during incremental learning sessions. The frozen simplex-ETF classifier geometry ensures consistent angular separation between classes across sequential training phases.

For the broader machine learning community, this work demonstrates that domain-specific challenges in remote sensing imagery can be effectively addressed through geometric constraints and cross-domain transfer learning. The methodology could extend beyond SAR applications to other sensor modalities facing similar data scarcity issues. The benchmark evaluation across 24 target classes organized into base plus seven incremental sessions provides realistic assessment of real-world performance degradation patterns.

The practical implications suggest that organizations deploying SAR-based automated systems can achieve significantly better accuracy with limited new training data by leveraging optical reference datasets. This approach reduces the annotation burden for incremental system updates while maintaining consistent performance across operational scenarios.

• →Optical-guided neural collapse framework transfers geometric structure from optical to SAR imagery for improved few-shot learning
• →Method achieves highest final accuracy and favorable trade-off between performance and degradation compared to recent FSCIL baselines
• →Principal angle constraints and frozen simplex-ETF geometry jointly induce neural collapse for better feature organization
• →Cross-domain transfer approach effectively addresses SAR-specific challenges of azimuth sensitivity and data scarcity
• →Framework demonstrates improved intra-class compactness and inter-class separability on 24-class benchmark evaluation