A neural network-based method for satellite-based mapping of sediment-laden sea ice in the Arctic
A neural network-based method for satellite-based mapping of sediment-laden sea ice in the Arctic
复制标题
DOI：
10.1016/j.rse.2021.112861
复制
发表时间：
2022-03
影响因子：
13.5
通讯作者：
Hisatomo Waga;H. Eicken;B. Light;Y. Fukamachi
中科院分区：
文献类型：
--
作者：
Hisatomo Waga;H. Eicken;B. Light;Y. Fukamachi
在北冰洋及其边缘海，海冰沉积物是一种普遍存在的现象。这项研究提出了一种基于卫星的方法，在量化的分布，允许更广泛的观测，在时间和空间监测时空变化的沉积物载冰。一个结构-光学模型与一个四流多层离散坐标方法辐射传输模型相结合，用于研究四种表面类型的表面光谱散射：干净的冰，沉积物负载冰与15种不同的沉积物负载从25到1000 g m−3，积水冰，和无冰开放水域。基于含沙冰与其他地表类型的光谱特征不同的事实，利用遥感地表反射率，采用最小二乘法的光谱分解算法，估算了含沙冰的比例。敏感性分析表明，50和500 g m− 3的沉积物负荷组合有效地代表了具有广泛沉积物负荷的含沉积物冰的面积分数。每种表面类型的估计分数和相应的遥感表面反射率用于训练人工神经网络，以相对于最小二乘法加快处理速度。比较来自这两种方法的分数的含泥沙的冰的面积分数的含泥沙的冰产生了良好的协议，突出了处理大型数据集的神经网络的上级性能。虽然我们的方法包含潜在的不确定性与方法的局限性，时空变化的沉积物载冰表现出合理的协议与文献中的空间格局和季节变化的原位观测的沉积物载冰。系统的卫星监测的沉积物冰分布可以提供广泛的，持续的，具有成本效益的观察，以促进我们的了解的作用，沉积物冰在各种各样的研究领域，包括沉积物输运和地球化学循环。
Sediment-laden sea ice is a ubiquitous phenomenon in the Arctic Ocean and its marginal seas. This study presents a satellite-based approach at quantifying the distribution of sediment-laden ice that allows for more extensive observations in both time and space to monitor spatiotemporal variations in sediment-laden ice. A structural-optical model coupled with a four-stream multilayer discrete ordinates method radiative transfer model was used to examine surface spectral albedo for four surface types: clean ice, sediment-laden ice with 15 different sediment loadings from 25 to 1000 g m−3, ponded ice, and ice-free open water. Based on the fact that the spectral characteristics of sediment-laden ice differ from those other surface types, fractions of sediment-laden ice were estimated from the remotely-sensed surface reflectance by a spectral unmixing algorithm using a least square method. Sensitivity analyses demonstrated that a combination of sediment loads of 50 and 500 g m−3effectively represents the areal fraction of sediment-laden ice with a wide range of sediment loads. The estimated fractions of each surface type and corresponding remotely-sensed surface reflectances were used to train an artificial neural network to speed up processing relative to the least squares method. Comparing the fractions of sediment-laden ice derived from these two approaches yielded good agreements for areal fractions of sediment-laden ice, highlighting the superior performance of the neural network for processing large datasets. Although our approach contains potential uncertainties associated with methodological limitations, spatiotemporal variations in sediment-laden ice exhibited reasonable agreement with spatial patterns and seasonal variations reported in the literature on in situ observations of sediment-laden ice. Systematic satellite-based monitoring of sediment-laden ice distribution can provide extensive, sustained, and cost-effective observations to foster our understanding of the role of sediment-laden ice in a wide variety of research fields including sediment transport and biogeochemical cycling.