Deep Neural Network (DNN) Based Synthetic Aperture Radar (SAR) Processor

Abstract:
Synthetic Aperture Radar (SAR) is one of the main sources of remote sensing data today. SAR raw data focussing is complicated and time consuming, therefore, is mostly done offline with sophisticated algorithms. Deep Learning (DL) has widespread use in the SAR applications domain but processing radar / SAR raw data into processed outputs has rarely been targeted. We have explored the potential of Deep Neural Networks (DNNs) to focus SAR raw data into Single Look Complex (SLC) images. Our method has potential for real-time processing of SAR data onboard airborne and spaceborne platforms. Promising results have been achieved for SAR raw data from the European Remote Sensing (ERS-1) satellite, originally acquired by the Alaska Satellite Facility (ASF). A Complex Valued DNN (CV-DNN) was designed and trained on two images of ~5700x27000 pixels over North America, which were broken down into smaller chips of 512x1024 pixels for training and validation purposes. The trained network was then able to focus raw data of a third image under test. A Structural Similarity Index (SSIM) of 0.749 with a mean squared error of 0.0029 was achieved for the output in comparison with the ground truth. Our research can serve as a steppingstone to exploit the unexplored potential of DNNs in the SAR focussing domain.