A novel Machine Learning based approach for Detecting Radiometric Artifacts in SAR Imagery

Abstract:
Synthetic Aperture Radar (SAR) images often suffer from radiometric artifacts originating due to many reasons leading to degradation of data product quality. To provide SAR data products with better quality, reliability and integrity, it is highly required to comprehend and detect the radiometric artifacts in SAR images based on its characteristic signature (either in time domain or frequency domain) using suitable approach so as to localize the artifacts for further detection. This paper describes a novel approach suitable for identifying these specific radiometric artifacts in SAR data based on AI-ML & Computer Vision techniques. Localized & global artifacts are dealt with different approaches, feature extraction using hand-crafted feature extractors & deep learning-based extractors are used. Training of classical Machine Learning models on extracted features as well as use of vanilla U-Nets & CNN for segmentation and classification is carried out. The suitability of different methods to increase the detection of varied artifacts is presented. Intersection Over Union (IoU) scores of segmentation based approaches & Recall – Precision score of classification based approach is compared for combination of various models & artifacts. The U-Net is able to handle most of the artifacts with ~0.85 IoU score & the ANN supersedes the other classifiers.