CauseHSI: Counterfactual-Augmented Domain Generalization for Hyperspectral Image Classification via Causal...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning
• Causal Inference

Background:

• Cross-scene hyperspectral image (HSI) classification is challenging due to domain shifts between source and target scenes.
• Existing single-source domain generalization (DG) methods struggle to adapt to unseen remote sensing environments.
• Causal theory provides a framework to understand distribution shifts and confounding biases in HSI data.

Purpose of the Study:

• To develop a causality-inspired framework (CauseHSI) for robust cross-scene HSI classification.
• To address interventional distribution shifts and confounding biases in HSI domain generalization.
• To improve the generalization capability of HSI classification models to unseen target scenes.

Main Methods:

• Proposed CauseHSI, a framework integrating causal theory into HSI domain generalization.
• Implemented a Counterfactual Generation Module (CGM) to simulate cross-domain interventions and generate diverse variants.
• Developed a Causal Disentanglement Module (CDM) to separate invariant causal features from spurious correlations using structural causal models.

Main Results:

• CauseHSI effectively simulates cross-domain interventions while maintaining semantic consistency.
• The CDM successfully disentangles domain-invariant causal semantics from domain-specific spurious correlations.
• CauseHSI demonstrates superior performance over existing DG methods on Pavia, Houston, and HyRANK datasets.

Conclusions:

• Aligning model learning with causal principles enhances robustness against domain shifts in HSI classification.
• CauseHSI offers a novel approach to address the fundamental obstacles in cross-scene HSI domain generalization.
• The proposed framework significantly improves the generalizability of HSI classification models for remote sensing applications.