Neural network retrieval of cloud parameters of inhomogeneous and fractional clouds: Feasibility study

One of the major issues of cloud parameter retrieval is how to optimize the improved observational capability of new radiometers to retrieve additional information about cloud characteristics. To investigate this problem, we developed a neural network approach for simultaneous retrieval of cloud parameters of inhomogeneous clouds with fractional cloud cover. We defined a simple inverse inhomogeneous cloud model with four parameters: mean optical depth, effective radius, relative cloud inhomogeneity, and fractional cloud cover. The retrieval algorithm, based on the use of a mapping neural network (MNN) with two hidden layers, was implemented and tested with synthetic multispectral reflectance data prepared for 2D clouds generated with a modified bounded cascade cloud model. We found that these cloud parameters could be retrieved from the moderate-resolution multispectral reflectance data with reasonable accuracy: for example with our data base, optical depth has a root mean square error of 1.7 for an ensemble of 1000 samples with optical depth up to 30. However, this accuracy depends on measurements errors and noises. A comparison with plan-parallel hypothesis shows the expected improvement of such an inhomogeneous cloud model. We show that the relevance of these cloud parameters is a function of the horizontal scale of averaging due to the net horizontal photon transport to and from adjacent cloud pixels. We tested the inclusion of ancillary data (reflectance of the neighbouring pixels) into the retrieval algorithm, and showed that the use of these ancillary data could partially correct the modeling error and significantly improve the performance of cloud parameter retrieval.