PSMNet: A Neural Network-Driven Approach for Pixel Similarity Measurement in Distributed Scatterer Interferometry

Pixel similarity measurement is a critical step in distributed scatterer (DS) interferometry, directly affecting DS phase estimation. Despite considerable efforts to improve its accuracy, existing methods still suffer from unsatisfactory performance, especially with small stack sizes. In recent years, deep neural networks have achieved remarkable breakthroughs in interferometric synthetic aperture radar (InSAR) processing. However, their potential for measuring pixel similarity in multitemporal InSAR remains unexplored. This article proposes a neural network-driven pixel similarity measurement approach, termed PSMNet. To address the challenge of accurately defining true data, a supervised learning strategy is designed. The proposed network consists of two main modules: 1) a feature extraction module that generates high-level feature images with enhanced representation and reduced noise and 2) a similarity measurement module that evaluates pixel similarity without relying on assumptions about data distribution. The network is trained on synthetic data, enabling it to generalize for different stack sizes and target characteristics. Extensive experiments on simulated and real TanDEM-X images demonstrate a significant accuracy improvement of the proposed approach, highlighting its robust performance for varying stack sizes and computational efficiency advantage compared to traditional methods. The proposed approach further enhances DS phase estimation and increases the number of measurement points, showing great promise for ground surface deformation monitoring.