Neural network retrieval of cloud parameters of inhomogeneous clouds from multispectral and multiscale radiance data: Feasibility study

In this paper, we investigated the feasibility of retrieving cloud parameters of inhomogeneous and fractional clouds from simulated multispectral and multiscale radiometric data by using mapping neural networks. A radiometric database prepared for neural network training consists of area‐averaged ra...

Full description

Saved in:
Bibliographic Details
Published in:Journal of Geophysical Research 2004-06, Vol.109 (D12), p.D12203-n/a
Main Authors: Cornet, Céline, Isaka, Harumi, Guillemet, Bernard, Szczap, Frédéric
Format: Article
Language:English
Subjects:
Atmospheric Composition and Structure
Cloud physics and chemistry
cloud retrieval
inhomogeneity
Inverse theory
Mathematical Geophysics
Meteorology and Atmospheric Dynamics
neural network
Ocean, Atmosphere
Radiative processes
Remote sensing
Sciences of the Universe
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, we investigated the feasibility of retrieving cloud parameters of inhomogeneous and fractional clouds from simulated multispectral and multiscale radiometric data by using mapping neural networks. A radiometric database prepared for neural network training consists of area‐averaged radiance data for two pixel scales, i.e., (1 km × 1 km) and (0.25 km × 0.25 km) pixels, respectively. The cloud parameter retrieval assumes a vertically uniform inhomogeneous and fractional cloud defined with 6 parameters, i.e., the mean and standard deviation of optical thickness, the mean and standard deviation of effective radius, the fractional cloud cover, and the cloud top temperature, all defined at a scale of cloud parameter retrieval. The retrieval procedure comprises two separate steps: the first one is relative to the angular interpolation and correction of radiance data (surface reflection and thermal emission contribution). The second step concerns the cloud parameter retrieval as such from interpolated and corrected radiance data. The input vector to the retrieval MNNs consists of 8 radiometric data in addition to a number of necessary ancillary data such as surface temperature and ground albedo. The 8 radiometric data are 5 area‐averaged radiances over (1 km × 1 km) pixel and 3 standard deviations of radiance over (1 km × 1 km) pixel estimated from (0.25 km × 0.25 km) pixel radiances. After evaluating the performance of the neural networks trained for each step, we tested the whole retrieval procedure for three types of inhomogeneous and fractional clouds: flat‐top bounded cascade clouds and flat‐top and non‐flat‐top Gaussian process clouds. All the cloud parameters of these clouds can be retrieved with reasonable accuracy in spite of the fact that the mean and standard deviation of optical thickness of non‐flat‐top clouds exhibit some dispersion. The inclusion of (0.25 km × 0.25 km) pixel radiance data as input vector components improved significantly the performance of the cloud parameter retrieval. Finally, we analyzed the consequences of some simplifying assumptions on the retrieved cloud parameters, and discussed the perspectives of the cloud parameter retrieval based on the neural networks.
ISSN:0148-0227
2156-2202
DOI:10.1029/2003JD004186