SE-CAE: A self-evolving CAE framework for noise-free InSAR

Interferometric synthetic aperture radar (InSAR) is a pivotal geodetic tool for deformation monitoring, yet multisource noise coupling in time-series data remains a critical challenge. This study proposes a self-evolving convolutional autoencoder (SE-CAE) framework to achieve noise-free deformation reconstruction. Validated using Sentinel-1 time-series data over Kunming, China, SE-CAE outperformed uncorrected data, spatiotemporal filtering, and GACOS-based methods through multimetric analysis. Quantitative evaluations demonstrated its superiority: an 82.01% reduction in standard deviation (STD), a 92.5% reduction in STD of displacements (SDD), and 88.66% improvement in spatiotemporal consistency (STC), alongside an 84.31% enhancement in local STC, confirming robustness across regional and localized scales. Spectral analysis verified its capacity to eliminate high-frequency noise while preserving deformation signals. Principal component analysis (PCA) revealed that SE-CAE addresses atmospheric delays (68.10% explained variance ratio, EVR), orbital errors (22.74% EVR), and topographic artifacts (2.80% EVR), accounting for 93.64% of corrected errors. Cross-validation with GNSS measurements further substantiated its reliability. By integrating deep learning with mechanistic interpretability, SE-CAE advances InSAR processing with balanced noise suppression, signal fidelity, and physically meaningful error correction, offering a transformative solution for high-precision deformation monitoring.