A-Benchmark-Dataset-for-Deep-Learning-based-InSAR-Phase-Unwrapping
We present a large and diverse benchmark dataset consisting of 31,100 pairs of wrapped and absolute phase image patches to support deep learning (DL) studies of phase unwrapping (PU) techniques in synthetic aperture radar interferometry (InSAR) signal processing. This dataset is abbreviated as InSAR-DLPU. For training and validation, we crop 256 * 256 pixel regions out of 1,000 absolute phase images (3,000 * 3,000 pixels) of real-world terrain data covering all of China, generated by the SRTM DEM database, creating 30,000 simulated pairs of wrapped and absolute phase image patches for training and 1,000 pairs for testing. In addition, 100 pairs of real TanDEM-X interferograms and their absolute phase image patches are provided to examine the generalization ability of DL-based PU methods. Based on this dataset, we can extensively validate and compare existing deep convolutional neural networks (DCNNs). We have made InSAR-DLPU publicly available, offering an important resource to support studies on DL-based PU in InSAR applications (e.g., topographic mapping and deformation monitoring).
The complete dataset of training and testing pairs is conveniently archived in MATLAB-compatible .mat file format. To better illustrate the above data, each pair of data is stored as color-rendered PNG images for quick look. The structure of InSAR-DLPU is as follows:
train_wrapped: The wrapped phase as input of the training dataset is in this folder and named 000001.mat to 030000.mat;
train_absolute: The absolute phase as ground truth of the training dataset is in this folder and named 000001.mat to 030000.mat;
test_wrapped: The wrapped phase as input of the testing dataset is in this folder and named 000001.mat to 001000.mat;
test_absolute: The absolute phase as ground truth of the testing dataset is in this folder and named 000001.mat to 001000.mat;
test_wrapped_real: The wrapped phase as input of the TanDEM-X testing dataset is in this folder and named 000001.mat to 000100.mat;
test_absolute_real: The absolute phase as ground truth of the TanDEM-X testing dataset is in this folder and named 000001.mat to 000100.mat.
The dataset is saved in Baidu Netdisk (https://pan.baidu.com/s/1UHLmMfnm55gH5B4YeWgtJg) and the access code is oej2.