基于Sentinel-1影像特征匹配的北极海冰漂移矢量提取

Sea ice drift vectors extraction based on feature tracking to Sentinel-1 images

海冰漂移对北极气候研究和人类活动保障有重要意义。针对星载辐射计或散射计数据采用模板匹配法提取的海冰漂移矢量结果存在空间分辨率较低及精度不佳等问题,本文以Sentinel-1影像为数据源,比较了4种常用的特征匹配算子SIFT、SURF、ORB、A-AKZE提取北极海冰漂移矢量的效果,并分析了HH与HV极化通道提取的漂移矢量提取的空间分布差异。针对特征匹配中不可避免的错误,结合既有算法的优势,本文提出了一套高效且准确的错误矢量滤除算法。通过MOSAiC浮标定位数据验证了本文方法提取的海冰漂移矢量的精度,并与基于Sentinel-1影像既有海冰漂移产品进行了精度对比。实验结果表明基于A-KAZE算子的算法在提取结果的数量与分布上优于SIFT、SURF与ORB算子;HH与HV极化通道影像提取的海冰漂移矢量在数量与空间分布上有所差异,结合两者可有效扩大海冰监测范围;通过错误矢量滤除算法能高效滤除错误匹配的同时保留更多正确漂移矢量。基于A-KAZE算子提取的海冰漂移矢量平均速度误差低于0.2 km/d,平均方向误差低于1°,与同样基于Sentinel-1 SAR影像但采用模板匹配法的DTU海冰漂移产品具有较高的一致性,但本方法提取的海冰漂移矢量具有更高的空间分辨率及更大的覆盖范围。

Sea ice drift is an important natural phenomenon in the Arctic, and it is important for climate research and human activities such as shipping security in the Arctic area. At present, sea ice drift products are often derived with space-borne radiometers and scatterometers with the template matching algorithm and suffer from low resolution and low accuracy. Sentinel-1 synthetic aperture radar imagery has high spatial resolution and holds great potential for deriving sea ice drift fields with high resolution and high accuracy by applying feature matching algorithms.This research compared sea ice drift results derived from four popular features including SIFT, SURF, ORB, and A-KAZE by using two pairs of Sentinel-1 Arctic sea ice SAR images. The similarities and differences between the performances of HH and HV imagery were also analyzed in terms of spatial distribution and coverage of the derived sea ice drift vectors. We proposed a filtering method combined with two published methods to identify incorrect vectors after the NNDR test with high calculation efficiency and accuracy. Finally, we evaluated the accuracy of sea ice drift vectors by comparing our derived results and DTU sea ice products with GPS data of MOSAiC buoys.Employing A-KAZE features to Sentinel-1 EW imagery can effectively derive sea ice drift fields with high spatial resolution and coverage rates. A-KAZE feature performs better than SIFT, SURF, and ORB in terms of spatial distribution and the number of vectors.Combining the vectors obtained from HH and HV polarization imagery can effectively extend the coverage of sea ice motion fields. The incorrect vector filter checks the similarity of a vector to its neighbors only if its speed or direction exceeds two times the standard deviation.It improves computational efficiency and retains more correct vectors than the two traditional methods. Validation with data of MOSAiC buoys found that the average speed error of sea ice drift vectors extracted using the proposed A-KAZE-based method was less than 0.2 km/d,and the average direction error was less than 1°. These products share a high consistency with DTU sea ice drift products obtained through employing Sentinel-1 SAR imagery but applying the template matching algorithm. However, our proposed methods presented a higher spatial coverage.This study demonstrates the potential of deriving sea ice drift vectors by applying dual-polarized Sentinel-1 SAR imagery and A-KAZE features. This approach can effectively and quickly generate sea ice drift vector fields of high spatial resolution with high spatial covering rates and high accuracy, which can serve as an accurate data source for climate research and maritime security in the Arctic.