Semi-Supervised Learning-Based Permanent Scatterer Selection Method for GB-InSAR

Permanent scatterer (PS) selection is a critical step in ground-based interferometric synthetic aperture radar (GB-InSAR) measurements. Conventional methods, often constrained by empirical thresholds, demonstrate limited capability in selecting PSs within low-coherence regions caused by deformations. Deep learning provides a data-driven alternative to characterize PS features. However, PSs lack an absolute truth standard. Current deep learning-based PS selection methods predominantly employ fully supervised learning approaches, where significant label noise exists in the training sets. This interferes with the feature learning process of networks and consequently reduces PS recall rates. To overcome these limitations, this article proposes a semi-supervised learning-based PS selection method for GB-InSAR. By establishing a closed-loop optimization path of 'refining the labels of training datasets and adjusting the strategy of model training,' this method suppresses the negative impact of label noise and enables the network to accurately fit the PS features, thus enhancing selection performance. First, initial reliable labels are generated via dual-threshold amplitude dispersion methods. Then, unlabeled pixels are processed through an optimal transport formulation solved via Sinkhorn-Knopp algorithm to improve label completeness. Finally, a mean teacher framework is introduced for semi-supervised training of the PS selection network, to systematically leverage implicit features of unlabeled data. A 3-D U-Net serves as the base network for pseudo-label generation and semi-supervised training, which processes sequential interferograms to output PS selection results. Simulated and measured datasets demonstrate the method's superiority over fully supervised and conventional methods, which significantly increases PS density and quantity while ensuring quality, and it particularly exhibits superior performance in deformation regions.