A novel lightweight 3D CNN for accurate deformation time series retrieval in MT-InSAR

Abstract

Multi-temporal interferometric synthetic aperture radar (MT-InSAR) is a powerful geodetic technique for detecting and monitoring ground deformation over extensive areas. The accuracy of these measurements is critically dependent on effectively separating unwanted phase signals, such as atmospheric delay effects (APS) and decorrelation noise. Recent advancements in data-driven deep learning (DL) methods have shown promise in phase separation by utilizing inherent phase relationships. However, the complex spatiotemporal relationship of InSAR phase components presents challenges that traditional 1D or 2D DL models cannot effectively address, leading to potential biases in deformation measurements. To address this limitation, we propose UNet-3D, a novel three-dimensional encoder-decoder architecture that captures the spatiotemporal features of phase components through an enhanced 3D convolutional neural network (CNN) ensemble, enabling accurate separation of deformation time series. In addition, a spatiotemporal mask is designed to reconstruct missing time series data caused by decorrelation effects. We also developed a separable convolution operator to reduce the computational costs without compromising performance. The proposed model is trained on simulated datasets and benchmarked against existing DL models, achieving an improvement of 25.0% in MSE, 1.8% in SSIM, and 0.2% in SNR. Notably, the computation cost is reduced by up to 80% through separable convolution, establishing the proposed model as both lightweight and efficient. Furthermore, a comprehensive analysis of performance factors was conducted to assess the robustness of UNet-3D, facilitating its open-source usability. To validate our approach in real-world scenarios, we conducted a comparative ground deformation monitoring study over Fernandina Volcano in the Galapagos Islands using Sentinel-1 SAR data and the Small Baseline Subset (SBAS) technique in MintPy software. The results show that the correlation between the deformation time series of UNet-3D and the SBAS method is as high as 0.91 and shows the advantages in mitigating the topography-related APS effects. Overall, the UNet-3D model represents a significant advancement in automating InSAR data processing and enhancing the accuracy of deformation time series retrieval.