Feasibility study of multi-pixel retrieval of optical thickness and droplet effective radius of inhomogeneous clouds using deep learning
Feasibility study of multi-pixel retrieval of optical thickness and droplet effective radius of inhomogeneous clouds using deep learning
复制标题
DOI：
10.5194/amt-10-4747-2017
复制
发表时间：
2017-12
影响因子：
3.8
通讯作者：
R. Okamura;H. Iwabuchi;K. S. Schmidt
中科院分区：
文献类型：
--
作者：
R. Okamura;H. Iwabuchi;K. S. Schmidt
抽象的。三维（3-D）辐射传输效应是卫星光学遥感云反演误差的主要来源。挑战在于，3D效果在多个卫星像素中表现出来，这是传统的单像素方法无法捕捉的。在这项研究中，我们提出了两种基于深度学习的多像素检索方法，这种技术在工程和其他领域的复杂问题中越来越成功。具体来说，我们使用深度神经网络（DNN）从多光谱、多像素辐射中获得云光学厚度和列平均云滴有效半径的多像素估计。第一种DNN方法使用相同两个波长的反射率基于平面平行均匀云假设校正传统的双谱反演。另一种DNN方法使用所谓的卷积层，并直接从四个波长的反射率中检索云的属性。DNN方法在大涡模拟的云场上进行训练和测试，这些云场用作3-D辐射传输模型的输入，以模拟向上的辐射。第二种基于DNN的检索，通过卷积层避开了双谱检索步骤，被证明是更准确的。它减少了3-D辐射传输的影响，否则会影响辐射值和估计云的属性鲁棒性，即使是光学厚云。
Abstract. Three-dimensional (3-D) radiative-transfer effects are a major source of retrieval errors in satellite-based optical remote sensing of clouds. The challenge is that 3-D effects manifest themselves across multiple satellite pixels, which traditional single-pixel approaches cannot capture. In this study, we present two multi-pixel retrieval approaches based on deep learning, a technique that is becoming increasingly successful for complex problems in engineering and other areas. Specifically, we use deep neural networks (DNNs) to obtain multi-pixel estimates of cloud optical thickness and column-mean cloud droplet effective radius from multispectral, multi-pixel radiances. The first DNN method corrects traditional bispectral retrievals based on the plane-parallel homogeneous cloud assumption using the reflectances at the same two wavelengths. The other DNN method uses so-called convolutional layers and retrieves cloud properties directly from the reflectances at four wavelengths. The DNN methods are trained and tested on cloud fields from large-eddy simulations used as input to a 3-D radiative-transfer model to simulate upward radiances. The second DNN-based retrieval, sidestepping the bispectral retrieval step through convolutional layers, is shown to be more accurate. It reduces 3-D radiative-transfer effects that would otherwise affect the radiance values and estimates cloud properties robustly even for optically thick clouds.