Estimation and analysis of geometric distortions in SAR images

Abstract

Synthetic Aperture Radar (SAR) is a kind of high-resolution imaging sensor with the advantages of all-day, all-weather. SAR-based analysis is used in many fields and plays an important role because of its characteristic of not being disturbed by weather and climate factors, among which the target detection task based on SAR is the key to many technologies. However, deep learning requires a large amount of sample data support, the high cost of obtaining highresolution SAR data sets brought certain difficulties to the research works in related fields. Image-to-image (I2I) translation has been extensively applied to natural images and the great success has attracted researchers in various research backgrounds. Generative Adversarial Network (GAN), one of the most successful I2I algorithms, has also brought changes to the data acquisition in remote sensing domain. By feeding an EO image into a well-trained model, a synthetic but realistic SAR images can be obtained. But one of the most critical problems between EO and SAR: geometric distortion, is still a big challenge for the EO-SAR translation task. This project aims to utilize the prior EO-SAR translation works to estimate and analyze the geometric distortions between EO and SAR images.