Welcome to PhUnBench.

2D phase unwrapping (PU) is a challenging problem that is integral to geospatial analysis. Recently, deep learning-based computer vision approaches to this problem have proliferated. Using two or more phase images of the same region, Interferometric Synthetic Aperture Radar (InSAR) can be used to measure minute changes in elevation (metres) or ground surface deformation (millimeters) over large areas (hundreds of kilometers). These measurements are critical for a number of areas of study, including the creation of topographic models, monitoring and assessing natural disasters such as earthquakes, landslides and volcanoes, glacial flow, and structural stability analysis for engineering applications. However, due to limitations in the phase image acquisition process, the phase unwrapping necessary to derive measurements used in these applications remains an extremely difficult problem. Although numerous techniques have been proposed and machine learning approaches have proliferated, there exists no benchmark to compare the quality and reliability of phase unwrapping algorithms. Accordingly, we construct PhUnBench, the first phase unwrapping benchmark. PhUnBench contains twenty-five 12,000px by 20,000px open source images acquired from the European Space Agency’s Sentinel-1 program along with ground truth. It allows deep learning models to be trained and compared against each other and against analytical models. We present initial baseline results using recent deep learning methods, along with a baseline from SNAPHU, a current standard in analytical phase unwrapping.

The dataset can be downloaded below: