Improving discrimination between clouds and optically thick aerosol plumes in geostationary satellite data
Improving discrimination between clouds and optically thick aerosol plumes in geostationary satellite data
复制标题
DOI：
10.5194/amt-2021-422
复制
发表时间：
2022-01
影响因子：
3.8
通讯作者：
D. Robbins;C. Poulsen;S. Siems;S. Proud
中科院分区：
文献类型：
--
作者：
D. Robbins;C. Poulsen;S. Siems;S. Proud
抽象的。云遮蔽是从卫星数据中检索地球物理特性的关键初始步骤。尽管经过几十年的研究，云的过检测和欠检测问题仍然存在。高含量的气溶胶，特别是来自沙尘暴或火灾的气溶胶，通常被归类为云，反之亦然。在本文中，我们提出了一个使用机器学习为Himawari-8上的高级Himawari成像仪（AHI）创建的云掩模。为了训练该算法，从AHI和云-气溶胶激光雷达与正交偏振（CALIOP）激光雷达数据创建了一个经视差校正的同位数据集。人工神经网络（ANN）的训练上的搭配数据，以识别云在AHI场景。由此产生的NN云掩模进行验证，并与日本气象协会（JMA）和气象局（BoM）生产的云掩模进行比较，用于许多不同的太阳能和观看几何形状，表面类型和气团。5个案例研究涵盖了一系列具有挑战性的情况下，云面具也展示了掩蔽算法的性能。对于所有类别的等效真阳性率（TPR），NN掩码显示出较低的假阳性率（FPR），在等效TPR值下，NN和JMA掩码的FPR分别为0.106和0.198，NN和BoM掩码的FPR分别为0.314和0.464。这表明NN掩码分别与JMA和BoM掩码相比，对于相同的命中率，可以准确识别1.11和1.28倍的非云像素。NN掩模在区分厚气溶胶羽流和云方面特别有效，这很可能是由于包含了0.47 μm和0.51 μm波段。在大多数情况下，NN云罩比当前的操作云罩有所改进，建议通过包含0.47 μm和0.51 μm波段来改进当前的操作云罩。并提供了数据，以促进未来的研究。
Abstract. Cloud masking is a key initial step in the retrieval of geophysical properties from satellite data. Despite decades of research problems still exists of over or under detection of cloud. High aerosol loadings, in particular from dust storms or fires, are often classified as cloud and vice versa. In this paper we present a cloud mask created using machine learning for the Advanced Himawari Imager (AHI) on board Himawari-8. In order to train the algorithm a parallax-corrected collocated data set was created from AHI and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) lidar data. Artificial neural networks (ANNs) were trained on the collocated data to identify clouds in AHI scenes. The resulting NN cloud masks are validated and compared to cloud masks produced by the Japanese Meteorological Association (JMA) and the Bureau of Meteorology (BoM) for a number of different solar and viewing geometries, surface types and airmasses. 5 case studies covering a range of challenging scenarios for cloud masks are also presented to demonstrate the performance of the masking algorithm. The NN mask shows a lower false positive rate (FPR) for an equivalent true positive rate (TPR) across all categories, with FPRs of 0.106 and 0.198 for the NN and JMA masks respectively and 0.314 and 0.464 for the NN and BoM masks respectively at equivalent TPR values. This indicates the NN mask accurately identifies 1.11 and 1.28 times as many non-cloud pixels, for the equivalent hit rate when compared to the JMA and BoM masks respectively. The NN mask was shown to be particularly effective in distinguishing thick aerosol plumes from cloud, most likely due to the inclusion of the 0.47 μm and 0.51 μm bands. The NN cloud mask shows an improvement over current operational cloud masks in most scenario and it is suggested that improvements to current operational cloud masks could be made by including the 0.47 μm and 0.51 μm bands. The collocated data are made available to facilitate future research.