A comparative survey on SAR image segmentation using deep learning

Abstract

Synthetic Aperture Radar (SAR) image is a radar system that observes topographic maps using microwaves as an active sensor. Due to the backscattering characteristics of SAR, speckle is distributed in the image, making it difficult to analyze. This paper investigates the classically used unsupervised method of SAR image segmentation that can easily recognize and analyze SAR images and the recently used deep learning algorithm, and compare the accuracy using performance metrics. Although the method using deep learning has the problem of insufficient dataset, it improves performance by 10-20% compared to unsupervised. Also, among deep learning algorithms, how the algorithms used in Electro Optical / Infrared (EO / IR) are used in SAR images and problems are investigated. In a recent study, the SAR image considered as a visible light image and applied it to a deep learning algorithm using eo to obtain results. In the future, more benchmark datasets for SAR images should be built, and research on deep learning algorithms using SAR data information will be conducted.