Objectives Interferometric synthetic aperture radar (InSAR) acquires ground elevation information and surface deformation information by interferometrically processing multiple synthetic aperture radar (SAR) complex images with certain temporal and spatial correlations. It has been widely applied in high-precision 3D mapping, as well as in the monitoring and early warning of geological disasters such as earthquake deformation, volcanic movement, and surface subsidence. Phase filtering is one of the key technologies in InSAR data processing, and the quality of interferograms significantly affects the accuracy of subsequent processing steps such as phase unwrapping and digital elevation model generation, but filtering effectiveness is often compromised by complex noise in the interferometric phase. The traditional phase filtering methods with manual parameter adjustment have not only low filtering accuracy but also low computational efficiency. Deep learning-based phase filtering methods have demonstrated great potential in automating this process and improving performance, but current convolutional neural networks (CNN) based methods overlook the non-local features causing limiting accuracy. Transformer-based networks, on the other hand, have strong global modeling capabilities and can extract non-local features from interference fringes, but they frequently struggle with capturing fine-grained local details and spatial structures of interference fringes, which are equally crucial for precise phase restoration.

Methods Deep learning is data-driven, yet there is currently no publicly available InSAR interferometric phase dataset. The proposed method generates a clean phase map using shuttle radar topography mission 30 m digital elevation model, and adds simulated noise to produce an interferometric phase map, thereby creating a dataset for training the network. Besides, to effectively improve the filtering effect, we propose a joint local and non-local phase filtering network combing CNN and Transformer (PFCT) for InSAR images. First, based on the strong global feature extraction capability of the robust Transformer network, the PFCT network structure is proposed for phase filtering. The network performs both local and nonlocal filtering at the same time, avoiding the accuracy limitation problem caused by the neglect of non-local features of interference fringes in existing CNN-based methods while also addressing the local detail deficiency typical of Transformer approaches. The architecture integrates CNN layers for capturing high-frequency details and spatial patterns with Transformer modules for modeling long-range dependencies across the fringe patterns. Then, a new complex loss function which combines point phase loss and phase gradient loss is proposed to guide the network to maintain the integrity of interference fringes while improving filtering performance.

Results In the simulation data experiment, the proposed method achieves a mean square error index that is 15.5% lower than the suboptimal algorithm, and the structural similarity (SSIM) index is 5.3% higher, indicating superior noise removal and structural preservation ability. These improvements are consistently observed across varying noise levels, confirming the model's robustness. The enhanced SSIM particularly underscores its effectiveness in maintaining critical fringe structures essential for accurate unwrapping. In the true data experiment, the residue removal index is 1.8% higher than the suboptimal method, demonstrating enhanced reliability in practical applications. Visual assessment of filtered interferograms also shows a marked reduction in speckle noise and better preservation of deformation fringes in areas with high phase gradients. The experiments show that the PFCT network model only slightly reduces computational efficiency, but achieves better filtering results than other methods, effectively maintaining the stripe structure while maintaining the filtering effect.

Conclusions The proposed method has a great filtering effect and powerful generalization ability on InSAR phase filtering across different terrains and noise conditions. By successfully hybridizing CNN and Transformer strengths, it provides a robust and efficient option for high-quality interferogram generation in modern InSAR processing workflows. Future work may focus on further optimizing the network for edge-device deployment and exploring its adaptation to multi-temporal InSAR time series analysis for dynamic monitoring applications.