Exploring Detection of Multilayer Clouds and Retrieval of Their Properties Using Multispectral Satellite Data in an Artificial Neural Network Approach

Satellite remote sensing is essential for monitoring clouds on regional and larger scales. Because clouds affect radiation, precipitation, atmospheric heating, and visibility, among other atmospheric variables, it is critical to understanding and modeling of those variables that cloud vertical structure be characterized as well as the bulk microphysics of the cloudy column. Most retrievals of cloud properties from satellites rely on interpretation of the observed radiances as emanating from a single, plane-parallel layer. A physical model is employed to convert the radiances into a several parameters, such as top height CTH, optical depth COD, and particle size CER, that are assumed to represent the average for the observed cloud. Yet, many cloud systems consist of one of more cloud layers, each having it own average properties. Determining the layers of the cloud system and their average properties is a daunting challenge for passive remote sensing. A variety of multispectral techniques have been employed to detect multilayer systems with varying degrees of success. Fewer methods have gone to the next step and attempt to retrieve the properties of the bottom and top layers. These various methods have been only mildly successful. Much of the difficulty in addressing this challenge has been the ambiguity of the signals used in the methods. Determining whether the observed scene is composed of a thin cirrus over a water cloud or thick cirrus contiguous with underlying layers of ice and water clouds is often difficult because of similarities in the observed radiance values. Recently, artificial neural network algorithms employing two or more infrared channels have been developed to detect thin and thick ice clouds and estimate CTH and COD regardless of the underlying background, clear or cloudy. They are based on training with CALIPSO and Cloudsat (CC) data matched with multispectral radiances from various satellites. The CC lidar and radar profiles provide the vertical structure, CODs, and CER values that serve as output for a neural network NN algorithm that employs various combinations of multispectral infrared radiances. By applying the trained NN to MODIS data, it is possible to estimate the height and optical properties of the upper-layer clouds. It may also be possible determine whether the cloud is mostly contiguous in the vertical or whether it comprises well-separated distinct layers. Knowing the properties of the upper part of the cloud and its layering status, it should be possible to determine the properties of the lower layer using a multilayer retrieval model. This paper explores the potential of this approach for better defining and monitoring multilayer cloud systems.