A Comparison of CFAR Object Detection Algorithms for Iceberg Identification in L- and C-band SAR Imagery of the Labrador Sea

Abstract

Abstract. In this study, we pursue two objectives: first, we compare six different “Constant-False-Alarm-Rate" (CFAR) algorithms for iceberg detection in SAR images, and second, we investigate the effect of radar frequency by comparing the detection performance at C- and L-band. The SAR images were acquired over the Labrador Sea under melting conditions. In an overlapping optical Sentinel-2 image, 492 icebergs were identified in the area. They were used for an assessment of the algorithms’ capabilities to accurately detect them in the SAR images and for the determination of the number of false alarms and missed detections. By testing the detectors at varying probability of false alarm (PFA) levels, the optimum PFA for each detector was found. Additionally, we considered the effect of iceberg sizes in relation to image resolution. The results showed that the overall highest accuracy was achieved by applying a Log-normal CFAR detector to the L-band image (F-score of 70.4 %), however, only for a narrow range of PFA values. Three of the tested detectors provided high F-scores above 60 % over a wider range of PFA values both at L- and C-band. Low F-scores were mainly caused by missed detections of small (< 60 m) and medium-sized (60–120 m) icebergs, with approximately 20–40 % of the medium icebergs and 85–90 % of small icebergs being missed by all detectors. The iDPolRad detector which is sensitive to volume scattering is less suitable under melting conditions.

Multi-Environment Adaptive Fast Constant False Alarm Detection Algorithm Optimization Strategy

“…In the non-uniform environment along the reference window before and after processing, when k units are deleted from one of the reference window edges by comparing the reference units with the mean variance, the corresponding function (MGF) of local estimation generated is calculated as follows [31]:…”