Cloud identification and property retrieval from Himawari-8 infrared measurements via a deep neural network
Cloud identification and property retrieval from Himawari-8 infrared measurements via a deep neural network
复制标题
DOI：
10.1016/j.rse.2022.113026
复制
发表时间：
2022-06
影响因子：
13.5
通讯作者：
Xinyue Wang;H. Iwabuchi;Takaya Yamashita
中科院分区：
文献类型：
--
作者：
Xinyue Wang;H. Iwabuchi;Takaya Yamashita
云是天气和气候系统的重要组成部分，而统一的云特性，如云顶高度(Cth)和云光学厚度(COT)，则需要提高精度和计算效率。建立了基于图像的深度神经网络(DNN)模型，用于喜马8号卫星红外测量中的云层识别和Cth和冰盖的同时反演。DNN模型以2016年4个月的亮温数据作为输入，以CloudSat和Cloud-Aerosol Lidar和红外探路者卫星观测(CALIPSO)的活跃遥感产品的云特性作为目标真实。添加了包括垂直温度分布、表面高程和几何参数在内的补充变量作为输入数据。首先用一个独立的数据集检验了DNN模式的性能，然后选择了一个CloudSat路径和一个Himawi-8粒子(85°E-205°E，60°S-60°N)上的个例，通过与两个物理模式的结果进行比较，进一步验证了该模式的有效性。对于水和冰-CTH的估计，DNN模型与目标值具有很高的一致性，对于COT≥为0.3的高冰云，总的CTH相关系数为0.9。值得注意的是，作为一种本质上的红外方法，DNN将可预测的冰盖扩展到~200，对于COT>；1的高冰云，DNN模型的相对偏差为~20%。DNN模型的强大精度主要来自于它对输入亮温图像上印记的空间特征进行学习的能力，以及它在三维空间中整合相邻像素的信息。使用DNN模式进行单次全盘估计需要一个处理器大约20分钟；因此，对于恶劣天气监测和中尺度云系研究，可以获得24小时内统一可用的近实时云属性反演。
Clouds constitute a key component of weather and climate systems, whereas the uniform retrieval of cloud properties, such as cloud top height (CTH) and cloud optical thickness (COT), requires accuracy and computational efficiency improvements. In this study, an image-based deep neural network (DNN) model for cloud identification and simultaneous retrieval of CTH and ice-COT is developed for Himawari-8 satellite infrared measurements. The DNN model is trained with brightness temperature data from four months in 2016 as the input, and cloud properties of an active remote sensing product from CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) as the target truth. Supplementary variables, including the vertical temperature profile, the surface elevation, and the geometrical parameters, are added as the input data. DNN model performance is first tested with an independent dataset, and then cases over a CloudSat track and a Himawari-8 granule (85°E–205°E, 60°S–60°N) are selected for further validation of the model by comparing its results with those from two physics-based models. For both the water- and ice-CTH estimates, the DNN model shows high consistency with the target values, with an overall CTH correlation coefficient of 0.90 for high ice clouds with COT ≥0.3. Notably, as an infrared method in nature, the DNN extends the predictable ice-COT to ~200, with relative biases of ~20% for high ice clouds with COT >1. The strong accuracy of the DNN model is primarily derived from its ability to learn from the spatial features imprinted on the input brightness temperature image, and its integration of information from neighboring pixels in a three-dimensional space. A single full disk estimation with the DNN model takes about 20 min using one processor; therefore, near-real-time cloud property retrieval that is uniformly available over 24 h can be obtained for severe weather monitoring and mesoscale cloud-system studies.