SPIE Journal Paper | 1 March 2010
KEYWORDS: Clouds, Neural networks, Scattering, Principal component analysis, Data modeling, Atmospheric modeling, Sun, Evolutionary algorithms, Carbon dioxide, Near infrared
In Part I a set of two layer feed-forward neural networks, trained via back propagation of sensitivities, was applied to a synthetic set of radiances in micro-windows of the near-infrared to make predictions of cloud water (cw), cloud ice (ci), effective scattering heights of cloud water and ice, (pcw and pci, respectively) and the column water vapor (w). A threshold test, using 2 g/m-2 for cloud water and 10 g/m-2 for cloud ice, was applied to the retrieved values to distinguish clear from cloudy scenes. In that work the discussion was limited to the nadir viewing geometry, and was applied only to land surfaces, excluding desert and snow and ice fields. Part II describes the extension to a set of high resolution radiances, as might be measured by a grating spectrometer from space, in both nadir and sun glint viewing geometries. Furthermore, results are given for all land surface types as well as scenes over ocean. Prior to neural network training, a Principal Component Analysis (PCA) is applied to the high resolution spectra, which consist of three bands centered at 0.76µm (O2 A-band), 1.61µm (weak CO2 band) and 2.06µm (strong CO2 band), each with 1016 channels. Analysis shows that the five leading EOFs together capture 99.9% of the variance in each band, reducing the data size by more than two orders of magnitude. Application of the trained neural networks to an independent data set, generated using CloudSat and Calipso cloud and aerosol profiles, as well as carbon dioxide profiles from a chemical transport model, were used to quantify the skill in the retrieval. The results vary significantly with surface type, viewing mode and cloud properties. Accuracies range from 7% to 100% (typically close to 75%), with confidence levels almost always greater than 90%.
