摘要:
合成孔径雷达干涉测量技术(Interferometric Synthetic Aperture Radar，InSAR)凭借全天时、高空间分辨率的特点在地表形变监测领域极具优势。然而，因采矿活动导致的大梯度沉降漏斗在干涉图中易产生密集的非线性干涉条纹，对相位解缠带来挑战。为解决这个难题，提出了一种将深度学习目标检测网络与沉降漏斗反演建模相结合的解缠方法，旨在识别和模拟干涉图中的沉降漏斗，在缠绕相位中降低相位梯度，从而实现高精度相位解缠。该方法首先在YOLOv10框架下采用模拟和真实矿区干涉图训练沉降漏斗探测模型，实现大规模干涉图沉降漏斗自动识别；然后，利用以角度偏差最小化为目标函数的二维高斯模型对每个沉降漏斗进行反演和建模；再次，从干涉图中减去模型相位，对残余相位滤波，减少因滤波导致的相位混叠并增加干涉信噪比；最后，对滤波的残余相位进行解缠并加回模型相位，恢复真实解缠相位。通过模拟数据和覆盖山西大同地区的陆探1号数据集，分析了新方法解缠前后相位残点的变化以及模拟数据统计，并与传统矿区解缠处理方法进行了对比。结果表明，新方法具有更强的鲁棒性和更高的精度，在矿区非线性高形变梯度区域的解缠中，平均均方误差较传统方法减小了57.4%， 平均残点数减小了23.4%，证明了所提出方法的可行性。
Abstract:
Objectives: Most phase unwrapping algorithms in InSAR technology rely on the assumption of phase continuity. However, mining subsidence is characterized by its small spatial extent, large-gradient subsidence, and sudden occurrences, which result in dense interference fringes in the interferograms, making it difficult to satisfy the phase continuity assumption. Methods: To address this challenge, proposes a phase unwrapping method that combines a deep learning-based object detection network with subsidence funnel inversion modeling. The goal is to identify and simulate subsidence funnels in the interferograms, reducing the phase gradient within the wrapped phase and thereby achieving high-precision phase unwrapping. The method first utilizes simulated and real-world interferograms from mining areas to train a subsidence funnel detection model within the YOLOv10 framework, enabling automatic large-scale identification of subsidence funnels in interferograms. Then, a two-dimensional Gaussian model, optimized using a minimum angular deviation objective function, is employed to invert and model each subsidence funnel. The modeled phase is subtracted from the interferogram, and the residual phase is filtered to reduce phase aliasing caused by filtering and to increase the signal-to-noise ratio of the interferogram. Finally, the filtered residual phase is unwrapped and added back to the modeled phase to restore the true unwrapped phase. Results: Using simulated data and the LuTan-1 dataset covering the Datong region in Shanxi, the proposed method was analyzed in terms of phase residual points before and after unwrapping, along with statistical results from the simulated data. Compared to traditional unwrapping methods used in mining areas, the proposed method demonstrates stronger robustness and higher accuracy. In nonlinear high-gradient deformation areas, the mean squared error was reduced by 57.4%, and the average number of residual points decreased by 23.4%. Conclusions: By combining model identification with fringe reduction, the wrapped phase better satisfies the phase continuity assumption, reducing the impact of large gradient deformations on subsequent unwrapping operations and improving the accuracy of the unwrapped results. Notably, the proposed algorithm can better restore the true phase at the center of subsidence funnels, a task that traditional 2D unwrapping algorithms struggle to achieve.
下载:
