Exploiting Deep Matching and SAR Data for the Geo-Localization Accuracy Improvement of Optical Satellite Images

Improving the geo-localization of optical satellite images is an important pre-processing step for many remote sensing tasks like scene monitoring over time or the scene analysis after sudden events. These tasks often require the fusion of geo-referenced and precisely co-registered multi-sensor data. Images captured by high resolution synthetic aperture radar (SAR) satellites have an absolute geo-location accuracy within few decimeters. This renders SAR images interesting as a source for the geo-location improvement of optical images, whose geo-location accuracy is in the range of some meters. In this paper, we are investigating a deep learning based approach for the geo-localization accuracy improvement of optical satellite images through SAR reference data. Image registration between SAR and optical satellite images requires few but accurate and reliable matching points. To derive such matching points a neural network based on a Siamese network architecture was trained to learn the two dimensional spatial shift between optical and SAR image patches. The neural network was trained over TerraSAR-X and PRISM image pairs covering greater urban areas spread over Europe. The results of the proposed method confirm that accurate and reliable matching points are generated with a higher matching accuracy and precision than state-of-the-art approaches