Multi-stage multi-task deep learning modelling for high-quality InSAR deformation analysis

Abstract

Spaceborne interferometric synthetic aperture radar (InSAR) has emerged as one of the most powerful and widely used geodetic tools for mapping ground deformation over large areas with high spatial resolution. Compared to traditional geodetic techniques such as leveling, which are labor-intensive and often restricted to smaller spatial extents, InSAR offers unparalleled advantages in terms of its broad coverage and capability to achieve sub-centimeter precision. By analyzing the interferometric phase difference between synthetic aperture radar (SAR) images acquired over the same area, InSAR is able to map surface deformation with remarkable accuracy. However, the interferometric phase includes not only the deformation signal of interest but also contributions from topography, atmospheric effects, and decorrelation noise. These non-deformation components pose significant challenges, as they require accurate compensation to ensure reliable deformation analysis. In particular, compensating for the atmospheric delays and noise-induced decorrelation using a single interferogram remains a limitation, leading to reduced accuracy in the derived deformation maps. Although several traditional techniques have been proposed to improve InSAR processing, these methods are computationally demanding, time-intensive, and often require expert knowledge for successful implementation. Given the exponential growth of InSAR data, there is an urgent need to develop more efficient processing approaches that minimize human intervention while maintaining high accuracy. Recent advancements in artificial intelligence (AI), particularly deep learning (DL), offer transformative opportunities for addressing these limitations in InSAR processing. Unlike traditional machine learning techniques that rely on predefined feature extraction, deep learning models leverage advanced architectures to learn directly from data, identifying complex patterns that are difficult to model explicitly. Popular DL architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), vision transformers (ViTs), and generative adversarial networks (GANs) have shown remarkable potential in processing large-scale datasets, making them highly suitable for InSAR applications. These models excel in both 2D spatial analysis and 1D temporal signal processing, which are essential for various InSAR tasks such as phase filtering, unwrapping, reconstruction, denoising, interpretation, and prediction. Multi-stage learning frameworks, which break complex tasks into smaller subtasks, and multitasking architectures, which share backbone structures to perform multiple tasks simultaneously, are particularly advantageous in enhancing performance. However, highly engineered DL models often come with significant computational costs, particularly for near-real-time InSAR applications on resource-limited hardware. These computational challenges remain a key obstacle to the widespread adoption of deep learning in InSAR processing workflows. Another critical limitation in applying DL to InSAR lies in the lack of sufficient ground truth data for training, particularly for tasks such as phase unwrapping, deformation retrieval, and noise reduction. To overcome this issue, researchers tend to develop sophisticated simulated datasets that replicate real-world scenarios, encompassing topographic effects, deformation signals, and atmospheric disturbances. These simulated datasets enable the development of task-specific DL architectures that are optimized for the unique characteristics of InSAR data. By combining diverse simulations with domain-specific models, DL approaches can better generalize across a wide range of InSAR challenges, significantly enhancing their robustness and reliability. This thesis addresses the applicability of deep learning in InSAR processing by presenting five significant contributions. First, to reconstruct masked phases caused by decorrelation, a multi-stage GAN architecture was proposed. This method divides the complex task of phase reconstruction into subtasks such as extracting fringe edges, reconnecting disconnected edges, and reconstructing the interferometric phase. By leveraging adversarial learning, the GAN effectively reconnects disconnected fringe edges and reconstructs the interferometric phase in a joint-learning framework, outperforming traditional reconstruction methods that rely solely on decorrelation masks. Second, a GAN-based model was developed to improve phase filtering and unwrapping, addressing the blurring artifacts typically observed in regular CNN models. By employing adversarial training between the generator and discriminator, this approach captures intricate phase patterns and reduces reconstruction artifacts, ensuring more accurate and realistic phase representations with different levels of decorrelation noise. Third, a robust 3D CNN model was designed to remove non-deformation components from InSAR time-series data by incorporating both spatial and temporal kernels. Unlike traditional methods that use fixed temporal windows, this model processes the entire InSAR data cube to avoid artifacts introduced by limited interferogram stacking. Additionally, computational costs were minimized through separable convolution operations, and a spatiotemporal mask was integrated to mitigate the effects of decorrelation. Fourth, a multitask ViT model was proposed for interpreting deformation in single interferograms. By sharing a common ViT backbone and branching into two output heads, this model simultaneously classifies deformation types and localizes deformation signals. The method is specifically designed to detect deformation signals, such as volcanic and coseismic events, even when blurred by non-deformation noise. The model utilizes transfer learning from natural image datasets to overcome the limited availability of labeled InSAR data, achieving high accuracy in both volcanic and coseismic deformation scenarios. Finally, the feasibility of the proposed methods was evaluated through extensive simulated and real-world experiments, demonstrating robust performance across diverse tasks and conditions. By advancing the use of deep learning architectures and leveraging realistic simulated datasets, this thesis contributes to bridging the critical gaps in InSAR processing, enabling faster, more accurate, and automated analysis.