Drifting icebergs represent a significant hazard for polar navigation and are able to impact the ocean environment around them. Freshwater flux and the associated cooling from melting icebergs can locally decrease salinity and temperature and thus affect ocean circulation, biological activity, sea ice, and –on larger spatial scales– the whole climate system. However, despite their potential impact, the large-scale operational monitoring of drifting icebergs in sea ice-covered regions is as of today typically restricted to giant icebergs, larger than 18.5 km in length. This is due to difficulties in accurately identifying and following the motion of much smaller features in the polar ocean from space. So far, tracking of smaller icebergs from satellite imagery thus has been limited to open-ocean regions not covered by sea ice. In this study, a novel automated iceberg tracking method, based on a machine learning-approach for automatic iceberg detection, is presented. To demonstrate the applicability of the method, a case study was performed for the Weddell Sea region, Antarctica, using 1213 Advanced Synthetic Aperture Radar (ASAR) satellite images acquired between 2002 and 2011. Overall, a subset of 414 icebergs (3134 re-detections in total) with surface areas between 3.4 km² and 3612 km² were investigated with respect to their prevalent drift patterns, size variability, and average disintegration. The majority of the tracked icebergs drifted between 1.3 km and 2679.2 km westward around the Antarctic continent, following the Antarctic Coastal Current (ACoC) and the Weddell Gyre, at an average drift speed of 3.6 ± 7.4 km day⁻¹. The method also allowed us to estimate an average daily disintegration (i.e. iceberg area decrease) rate of ~0.13% (~37% year⁻¹) for all icebergs. Using the sum of all detected individual surface area reductions, we estimate a total iceberg mass decrease of ~683 Gt year⁻¹, which can be freshwater input and/or new ‘child’ icebergs calved from larger icebergs. The extension to an automated long-term tracking method for icebergs is challenging as the iceberg shape can vary significantly due to abrupt disintegration or calving of bergy bits. However, our machine learning approach extended by automatic shape-based tracking capabilities proved to be a reliable alternative for automatic detection and tracking of icebergs, even under the ambiguous SAR background signatures often found in the Southern Ocean. In particular, the method works in the challenging near-coastal environment where the presence of sea ice and coastal ocean dynamics such as surface waves usually pose major obstacles for other approaches.

漂流的冰山对极地航行构成重大危害，并可能影响其周围的海洋环境。淡水通量和融化的冰山带来的相关冷却作用会局部降低盐分和温度，从而影响海洋循环，生物活动，海冰，以及在更大的空间范围内影响整个气候系统。但是，尽管有潜在的影响，但到今天为止，在大面积冰山中对海冰覆盖地区的流冰山进行大规模运行监控通常仅限于长度超过18.5公里的巨型冰山。这是由于难以从太空准确识别和跟踪极地海洋中较小特征的运动。到目前为止，从卫星图像中追踪较小的冰山仅限于没有海冰覆盖的开放海洋地区。在这个研究中，提出了一种新颖的基于机器学习方法的自动冰山跟踪方法，用于自动冰山检测。为了证明该方法的适用性，使用2002年至2011年间获得的1213枚先进合成孔径雷达（ASAR）卫星图像，对南极韦德尔海域进行了案例研究。总体而言，有414个冰山的子集（3134次重新检测）总面积）在3.4 km之间^研究了2和3612 km ²的普遍漂移模式，大小变异性和平均崩解。大部分被追踪的冰山跟随南极沿海海流（ACoC）和Weddell Gyre沿南极大陆向西漂移1.3 km至2679.2 km，平均日漂移速度为3.6±7.4 km ^-1。该方法也让我们估计平均每天解体（即冰山面积减少）〜0.13％率（〜37％，比去年^-1）的所有冰山。使用所有检测到的个体表面积减少的总和，我们估计冰山的总质量减少约683 Gt年^-1，可以是淡水输入和/或从较大的冰山中割下的新的“儿童”冰山。冰山的自动长期跟踪方法的扩展具有挑战性，因为冰山的形状可能会由于突发碎裂或切碎的碎裂而发生很大变化。但是，即使在南大洋上经常出现的模棱两可的SAR背景特征下，我们的机器学习方法也扩展了基于形状的自动跟踪功能，是自动检测和跟踪冰山的可靠选择。尤其是，该方法在具有挑战性的近海岸环境中起作用，在该环境中，海冰和诸如面波之类的沿海海洋动力学的存在通常对其他方法构成主要障碍。