New image analysis technologies for fast and accurate retrieval of sea ice floe size distribution (FSD) from satellite SAR imagery

北极海冰正在迅速变化。最深刻的例子发生在2012年夏天，这是自1979年卫星传感器开始监测海冰以来记录的最低冰面积。2012年8月，一场猛烈的风暴经过北极，短短一周内，几乎相当于两倍英国面积（40万平方公里）的海冰消失了，到8月中旬，北极西部的无冰水域高达北纬80度。导致海冰如此迅速减少的关键过程之一是海冰在冬夏过渡期间的破碎。在这个过渡期间，海冰的边缘向北退缩，暴露出更大的开阔水域，产生波浪，传播到浮冰中，使更大的海冰分裂成更小的浮冰。随着浮冰变小，它们融化得更快，变得更有活力。与此同时，更多的太阳辐射通过暴露的开阔水域被吸收，这使得上层海洋变暖，反过来促进了冰的更快融化。这种连锁反应可以加速海冰的退缩，从而影响最小冰面积。在海冰/气候模型中，这种重要的浮冰破裂及其相关影响没有得到很好的体现。这部分是由于缺乏对我们目前对过程的知识的理解和验证，以及由于过程的复杂性使得很难有效地将它们实现为模型中的“简单”表示。要产生有效的参数化，需要准确的现场流粒径分布（FSD）数据，这些数据可用于验证和完善已知的参数化，以及制定新的参数化。卫星合成孔径雷达（SAR）提供了不受黑暗或云层影响的海冰观测，因此提供了理想的原始数据，可以从北极黑暗的冬季到多云的夏季检索消防处。在北极获取的卫星SAR图像越来越多，图像的空间分辨率通常高达1-20米。更重要的是，卫星SAR图像正在通过自主浮标系统获取，并与实地活动相结合。这为测量海洋、海冰和大气参数以研究复杂的浮冰破碎过程提供了理想的框架。然而，我们面临的挑战是缺乏经过验证的高质量算法，可以快速准确地从卫星SAR图像中获得FSD。先前应用于该问题的阈值算法不足以进行定量分析，并且性能没有得到精确评估。在这个项目中，我们首次将海冰物理与边缘图像处理技术结合起来，在完全不同的水平上开发FSD算法。我们利用最新的图像处理技术，包括a)小波算法，以减少散斑噪声，同时增加冰与水之间的边界对比度；b)基于局部统计的算法，从背景开放水域提取浮冰特征；c)结合边缘保持分水岭和分裂合并算法，有效地分割浮冰的接触边界。我们期望这套新算法能够从卫星SAR图像中获得更精确的海冰FSD，并为开发可用于建立海冰FSD长期数据库的通用算法奠定基础。

Arctic sea ice is changing rapidly. The most profound example is during the summer of 2012, in which the lowest ice extent was recorded since satellite sensors began to monitor sea ice in 1979. Within just one week, as a violent storm passed the Arctic in August of 2012, sea ice area equivalent to nearly twice the size of UK (0.4 million square kilometres) disappeared, leaving ice-free water up to 80 N by mid-August in western Arctic. One of the key processes that cause such rapid sea ice decline is sea-ice floe breakup during the winter-to-summer transition. During this transition the edge of sea ice retreats to the north, exposing larger open water fetch to generate waves, propagating into the ice pack, allowing larger sea-ice floes to break into smaller ones. As the floes become smaller, they melt faster and become more dynamic. At the same time more solar radiation is absorbed through exposed open water areas, which makes the upper ocean layer warmer and in turn promotes faster ice melting. This chain reaction can accelerate the sea ice retreat and thus impact the minimum ice extent. This important floe breakup and associated effects are poorly implemented in sea-ice/climate models. This is partly due to lack of understanding and verification of our current knowledge on the processes as well as due to complexity of the processes that makes it difficult to effectively implement them into "simple" representations in the models. Producing effective parameterisations requires accurate data on in-situ floe size distribution (FSD) that can be used to verify and refine the known parameterisations as well as to formulate new ones. Satellite Synthetic Aperture Radar (SAR) provides observations of sea ice unhindered by either darkness or cloud, thus provide ideal raw data from which FSD can be retrieved from dark winter to cloudy summer in the Arctic. There is an increasing number of satellite SAR images being acquired in the Arctic, and often at spatial resolutions in the images as good as 1-20 m. More importantly satellite SAR images are being acquired over autonomous buoy systems and in conjunction with field campaigns. This provides the ideal framework to measure the full range of ocean, sea-ice and atmosphere parameters to investigate complex floe breakup process. However the challenge we have is a lack of proven-quality algorithms that can derive FSD from satellite SAR images fast and accurately. Thresholding algorithms previously applied to the problem are not adequate for quantitative analysis and the performance has not been precisely assessed. In this project we, for the first time, combine sea ice physics with edge-cutting image processing techniques to develop FSD algorithms at a completely different level. We leverage the latest image processing technologies which include a) wavelet algorithms to reduce the speckle noise while increasing the contrast of the boundary between ice and water, b) local-statistics based algorithm to extract ice floe features from the background open water, c) and a combination of edge-preserving watershed and split-and-merge algorithms to effectively split up the touching boundary of the floes. We expect this set of new algorithms will produce much more accurate FSD from satellite SAR images, and lay a foundation develop universal algorithm that can be used to build a long-term sea-ice FSD database.