PULSAR: A Modified Dataset for Machine Learning in Remote Sensing Radars Phase Unwrapping

The application of deep learning algorithms for solving phase unwrapping problems in radar remote sensing data processing is an important area of research for the last years. Convolutional neural networks after their appropriate training allow one to unwrap phase interferograms quickly and using low computational resources. In the present work, we propose the PULSAR (Phase Unwrapping for Learning in Synthetic Aperture Radar Systems) dataset for supervised training and evaluation of deep learning phase unwrapping models in interferometric synthesized aperture radars (InSAR) data processing. The dataset is based on 100 absolute phase real InSAR samples. For each absolute interferogram, we generated 100 different wrapped phase interferograms by variation of a phase slope, noise level, height ambiguity, and positions of discontinuity points. The total dataset size is 10000 interferograms of 512×512 elements. To demonstrate our dataset, we used it to train phase unwrapping models based on the U-Net and ResUNet architectures. Experiments with real satellite data have shown satisfactory results: the relative error for various control samples is lower than 20%.