Phase unwrapping remains a critical and inherently challenging task in interferometric synthetic aperture radar (InSAR) data processing, particularly for applications such as ground deformation monitoring, terrain mapping, and infrastructure stability assessment. Traditional and heuristic-based methods often exhibit limitations in the presence of high noise levels, phase discontinuities, and complex surface topographies, leading to compromised accuracy and reliability. In this study, we introduce VOH-Net (Vision-Optimized Hybrid Network), a novel deep learning architecture that synergistically combines convolutional neural networks (CNNs) with enhanced long short-term memory (E-LSTM) models within a unified encoder-decoder framework. The encoder extracts multi-scale spatial representations from wrapped phase maps, while the E-LSTM modules model contextual dependencies across spatial domains to facilitate more coherent phase reconstruction in the decoder. A hybrid loss function, comprising total variation, variance of error, and mean squared error, is employed to enforce smoothness, structural consistency, and pixel-wise precision. Comprehensive experiments conducted on both synthetic and real InSAR interferograms demonstrate that VOH-Net consistently outperforms conventional and state-of-the-art deep learning-based methods. Specifically, VOH-Net achieves a 21.3 % reduction in RMSE, an 18.7 % improvement in phase continuity index, and an overall unwrapping accuracy exceeding 94.6 % on challenging terrain scenarios. Moreover, it shows robust generalization capabilities across varying noise levels and interferometric conditions. These quantitative gains underscore the effectiveness of the proposed architecture in handling discontinuities and noise-corrupted inputs. VOH-Net thus offers a powerful, scalable solution for next generation InSAR applications, paving the way for more reliable and automated satellite-based geophysical monitoring.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号