DLHPS: A novel DS-InSAR Homogeneous Pixel Selection Method Based on Prior Constraints and Consistency Learning

Homogeneous pixel selection (HPS) is a critical component in distributed scatterer interferometric synthetic aperture radar (DS-InSAR) processing, and it directly affects the accuracy and stability of phase linking and deformation retrieval. Conventional HPS approaches mainly rely on statistical goodness-of-fit tests applied to amplitude time series (e.g., Kolmogorov–Smirnov (KS), Anderson–Darling (AD), and ttest) to determine homogeneity; however, they often suffer from insufficient detection in areas with limited image numbers or complex scattering mechanisms. In recent years, deep learning has been introduced into HPS to learn local scattering structures and spatial patterns, but existing strategies typically depend on manually labeled samples or use statistical-test outputs as pixel-wise pseudo-labels for all pixels within a window. Such designs are vulnerable to pseudo-label noise and severe class imbalance, causing conservative predictions, an insufficient number of homogeneous pixels, and unstable spatial patterns. To address these issues, we propose a prior-constrained and consistency-learning DS-InSAR homogeneous pixel selection method, termed DLHPS. DLHPS constructs a statistical prior by fusing voting results from KS, AD, and ttest with respect to the window-center reference pixel, and further extracts high confidence homogeneous and high confidence non-homogeneous sample sets. By replacing dense hard supervision over all window pixels with sparse high-confidence constraints, DLHPS alleviates imbalance-induced degradation and reduces the adverse impact of pseudo-label noise. In addition, DLHPS incorporates amplitude-perturbation-based data augmentation with a dual-view consistency constraint, together with a lightweight spatial coherence regularization, to improve robustness and spatial continuity. Experimental results demonstrate that DLHPS achieves a 90.55% increase in mean coherence and a 71.89% reduction in phase residuals, providing more reliable homogeneous neighborhoods for subsequent DS-InSAR phase linking.