ESCORT: Enriched SAR Change detection via Copula Regression Technique

Synthetic Aperture Radar (SAR) change detection is a key tool for monitoring land-use and land-cover dynamics under all-weather and day/night conditions, yet its performance is often hindered by speckle noise, weak exploitation of spatial context, and limited modeling of nonlinear temporal dependencies. These issues make it difficult to reliably separate change and no-change regions, particularly in heterogeneous and transitional areas. This paper presents an Enriched SAR Change detection via Copula Regression Technique (ESCORT), which adopts a nonparametric framework designed to address these challenges through dependency-aware and spatially consistent modeling. Particularly, the core contribution lies in the use of local copula density (LCD) features, which capture nonlinear temporal dependencies between bi-titemporal SAR observations in a distribution-free manner and provide strong discriminative power under noise. To further stabilize these dependencies, spatial weighting kernels (SWK) are employed to enhance both the difference image and the input images while preserving spatial structure. Finally, an enriched feature construction strategy is proposed to explicitly encode neighborhood-level information, by which coherent change maps can be produced using a regularized kernel logistic regression (KLR) model with spatial consistency constraints. Experimental results on three diverse SAR datasets demonstrate the effectiveness of the proposed framework, achieving the highest Kappa scores on two datasets with relative Kappa improvement of up to 1.85% (up to 13.77% statistical gain) over the best baseline on San Francisco and/or Sulzberger datasets, while remaining competitive with recent deep learning approaches on the more challenging Ottawa dataset.