MiKA-HAD: Minority-Augmented KAN-Based Hybrid Self-Supervised Framework for Hyperspectral Anomaly Detection

In recent years, autoencoder (AE)-based models have demonstrated significant potential in hyperspectral anomaly detection (HAD). However, the inherent class imbalance between anomaly and background samples, coupled with substantial intraclass heterogeneity within background components, often leads traditional models to excessively capture majority-class background features, thereby weakening their discriminative capabilities for minority-class backgrounds and anomaly targets. Concurrently, conventional methods fail to sufficiently model nonlinear features and adapt to complex environments, undermining their detection performance and robustness in intricate environmental interference. To address these challenges, we propose MiKA-HAD, a minority-augmented Kolmogorov-Arnold network (KAN)-based hybrid self-supervised framework for HAD. It incorporates a density-based clustering-guided sample balancing strategy that dynamically synthesizes minority background samples with spectral-spatial diversity through self-supervised learning. This approach achieves feature space rebalancing while effectively suppressing noise interference on anomaly boundaries. To overcome high-dimensional data complexity and nonlinear feature extraction challenges, we construct a hybrid self-supervised network with a KAN-based multipath collaborative module that orchestrates synergy between local sensitivity preservation, nonlinear feature modeling, and global consistency maintenance. This tripartite architecture establishes a dynamic equilibrium that enhances representation capability and environmental robustness. Extensive experiments show that MiKA-HAD achieves superior detection accuracy and stability compared to existing approaches, particularly in complex environments with varying noise conditions. The framework establishes a new paradigm for robust HAD by addressing both class imbalance bias and nonlinear feature modeling limitations.