Deep learning and physics-based approaches for ice-ocean monitoring

北极的海冰覆盖正在发生重大变化。冰层范围和厚度的减少，以及开放水域季节的增加，正在加大这一生态敏感地区的压力。世界各地的冰雪服务运营以及政府决策者都迫切需要监测这些变化的方法。海冰浓度是一个关键变量，它表示海洋中某一特定区域被冰覆盖的比例。通常使用在电磁频谱的低频微波部分获得的遥感数据对其进行监测，因为这些频率的数据对云层或阳光不敏感。海冰浓度监测采用被动和有源两种微波传感器。被动微波传感器测量地球自然发射的能量，而主动微波传感器向地球发送信号并测量后向散射。被动微波传感器用于监测大范围(例如30公里)的海冰浓度，而合成孔径雷达(SAR)传感器提供更高分辨率的数据(例如50米)。合成孔径雷达传感器非常适合监测边缘冰区，在边缘冰区，冰涡、浮冰和海浪显著调制着海冰覆盖。最近推出的多个SAR星座系统提供了来自SAR的快速增长的数据量。这随后激发了人们对从这些数据中提取信息的自动化方法的浓厚兴趣。深度学习是一种数据驱动的方法，可以从数据中学习模式，非常适合这项任务。然而，这些方法被广泛视为“黑匣子”，阻碍了深度学习在下游应用中的广泛接受。深度学习算法提供的错误信息可能会危及在冰上操作的人。例如，缺乏冰层加固的渔船需要准确的冰缘位置。虽然已经提出了几种“打开盒子”的方法，但这些方法是针对具有不同对象和颜色的日常图像的问题空间而设计的。合成孔径雷达海冰图像是多尺度现象的灰度图像。此外，海冰是一个可以用微分方程式模拟的物理系统。拟议的研究计划将利用这些方面来开发易处理的、规模感知的、物理指导的方法。与以往在这个问题领域的工作不同，本文将使用真实数据和模拟数据来开发深度学习方法。这将允许对拟议的方法进行彻底调查。预期成果是改进卫星传感器对边缘冰区海冰密度的估计，并以目前的主要产品为基准进行定量比较；为多尺度物理系统建立一个新的深度学习框架。

Sea ice cover in the Arctic is undergoing significant change. Decreasing ice extent and thickness, and an increasing open water season, are driving increased pressures on this ecologically sensitive region. Ways to monitor these changes are desperately needed by ice service operations worldwide as well as governmental decision makers. Sea ice concentration is a key variable that indicates the fraction of a specified area of the ocean that is covered by ice. It is typically monitored using remote sensing data acquired in the low-frequency microwave portion of the electromagnetic spectrum because data at these frequencies are not sensitive to cloud cover or sunlight. Both passive and active microwave sensors are used for sea ice concentration monitoring. Passive microwave sensors measure the energy naturally emitted by the earth, while active microwave sensors send a signal to the earth and measure the backscatter. Passive microwave sensors are used to monitor sea ice concentration at large scales (e.g., 30 km), while synthetic aperture radar (SAR) sensors provide higher resolution data (e.g. 50 m). SAR sensors are well suited to monitoring the marginal ice zone, a realm where ice eddies, floes and waves significantly modulate the sea ice cover. The recent launch of multiple SAR constellation systems is providing a rapidly growing data volume from SAR. This has subsequently spurred strong interest in automated methods to extract information from these data. Deep learning is a data-driven approach that can learn patterns from data, and is well suited to this task. However, these methods are widely viewed as a `black box', which hinders widespread acceptance of deep learning in downstream applications. Erroneous information provided by a deep learning algorithm could endanger those operating on the ice. For example, fishing boats that lack ice-strengthening need an accurate ice edge location. While there are several 'opening the box' methods proposed, these are designed for the problem space of everyday images that have distinct objects and colours. SAR sea ice images are greyscale images of multi-scale phenomena. Additionally, sea ice is a physical system that can be modelled using differential equations. The proposed research program will exploit these aspects to develop tractable, scale-aware, physically guided approaches. Both real data and simulated data will be used to develop the deep learning approaches, which is different from past work in this problem domain. This will allow a thorough investigation of the proposed methodologies. Expected outcomes are improved estimates of sea ice concentration in the marginal ice zone from satellite sensors benchmarked quantitatively against current leading products; and a novel deep learning framework for multiscale physical systems.