Switchable Copula-guided Transformer diffusion model for detecting small sea-surface targets

Abstract

Abstract Detecting low, slow and small maritime targets remains difficult because sea clutter is non-uniform, non-stationary, non-Gaussian and often observed at extremely low signal-to-clutter ratios. Here we present a switchable Copula-guided Transformer denoising diffusion probabilistic model (SCT-DDPM) for unsupervised radar target detection. The framework replaces the conventional i.i.d. Gaussian forward prior with Copula-structured noise, uses a copula-aware weighted MSE objective to align denoising with the observed dependence structure, and combines diffusion-based anomaly attention with Transformer encoding to score target-like deviations from normal clutter. Across 13 Copula families on the IPIX radar dataset, SCT-DDPM maintained high detection performance. These results show that dependence-aware diffusion priors can improve target-clutter separability and provide a statistically grounded route for configuring diffusion models in complex maritime environments.