Low-Intervention Boundary-Risk Graph Calibration for Cross-Domain Few-Shot Hyperspectral Image Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral sensors generate high-dimensional data crucial for land-cover analysis.
• Reliable classification is challenging with few available target-scene labels, especially in cross-domain few-shot scenarios.
• Sparse support samples lead to unstable classification decisions in low-margin regions.

Purpose of the Study:

• To propose a risk-controlled calibration framework, Boundary-Risk Graph Calibration (BRGC), to enhance decision reliability in few-shot hyperspectral image classification.
• To improve the interpretation of scarce-label hyperspectral imagery.
• To address classification instability in low-margin regions.

Main Methods:

• BRGC combines boundary-aware mixability training with inference-time graph residual calibration.
• During inference, support labels are clamped, and an unlabeled target-query graph provides structural smoothing.
• Classification scores are modified via low-margin gating and bounded residual updates.

Main Results:

• BRGC consistently improves base classifiers across 10 target datasets with 10 random seeds.
• Achieved the highest macro-average Overall Accuracy (OA), Average Accuracy (AA), and Kappa among transductive baselines.
• Demonstrated improved performance through repair/damage diagnostics, ablation studies, and parameter sensitivity analysis.

Conclusions:

• BRGC effectively converts query-graph structure into low-intervention reliability evidence for improved hyperspectral image interpretation.
• The framework enhances the reliability of support-set decisions in challenging few-shot learning scenarios.
• BRGC offers a robust solution for accurate land-cover analysis with limited labeled data.