METHOD FOR THE COMPUTER-IMPLEMENTED GENERATION OF A SYNTHETIC DATA SET FOR TRAINING A CONVOLUTIONAL NEURAL NETWORK FOR AN INTERFEROMETRIC SAR

摘要

The present invention describes a method for the computer-implemented generation of a synthetic data set for training a convolutional neural network for an Interferometric Synthetic Aperture Radar. In a first step, a number of interferometric phase images being generated from a respective digital elevation model (DEM) is retrieved, where each digital elevation model (DEM) is back-geocoded from an image taken with a SAR system using a known acquisition geometry. In a second step, a number of amplitude images and coherence images is generated by providing a set of different patterns derived from ramp profiles and from natural patterns and by mapping the natural pattern digital values into a scale used for the ramp profile digital values. In a last step, a number of noisy images is repeatedly generated by grouping one of the phase images, one of the amplitude images and one of the coherence images into a triple and combining each triple to form a respective noisy image, wherein each of the noisy images is assigned to one of a plurality of different categories.