Neural network retrieval of cloud parameters from high-resolution multispectral radiometric data: A feasibility study

Abstract

One of the important issues of the cloud parameter retrieval is how to optimize the improved observational capability of new radiometers. In this study, we examined a neural network approach to retrieve simultaneously optical depth and effective radius of overcast bounded cascade clouds from high-resolution multiwavelength radiometric data. The high-resolution retrieval allows the assumption of uniform cloud parameters within the target pixel but also requires the integration of radiometric data of neighboring pixels as ancillary data because of the net horizontal transport of photons. The performance of the mapping neural network (MNN) high-resolution retrieval was evaluated under conditions of vertical and oblique illumination using six pairs of wavelengths with 0.64, 1.6, 2.2, and 3.7 μm. Two types of clouds are used: inhomogeneous clouds with horizontal uniform effective radius and inhomogeneous clouds with horizontally variable optical depth and effective radius. The results show that we can retrieve these cloud parameters with a reasonable accuracy, which varies with the spectral channels used for the retrieval. The application of a “one-neuron” MNN for the cloud parameter retrieval shows that the effective radius estimation depends on visible wavelength when used with another having only a small absorption as 1.6 or 2.2 μm.