Neural Networks algorithm development for polarimetric observations of above cloud aerosols (ACA)

The direct and indirect radiative effects of above clouds aerosols (ACA) are still highly uncertain in current climate assessments. Much of this uncertainty is observational as most orbital remote sensing algorithms were not designed to simultaneously retrieve aerosol and cloud optical properties. Recently, several algorithms have been developed to infer ACA loading and properties using passive, single view angle instruments (OMI, MODIS). Yet, these are not operational and still require rigorous validation. Multiangle polarimetric instruments like POLDER, and RSP show promise for detection and quantification of ACA. However, the retrieval methods for polarimetric measurements entail some drawbacks such as assuming homogeneity of the underlying cloud field for POLDER and retrieved cloud effective radii as an input into RSP scheme. In addition, these methods require computationally expensive RT calculations, which precludes real-time polarimetric data analysis during field campaigns. Here we describe the development of a new algorithm to retrieve atmospheric aerosol and cloud optical properties from observations by polarimetrically sensitive instruments using Neural Networks (NN), which are computationally efficient and fast enough to produce results in the field. This algorithm is specific for ACA, and developed primarily to support the ORACLES (ObseRvations of Aerosols above CLouds and their intEractionS) campaign, which will acquire measurements of ACA in the South-East Atlantic Ocean during episodes of absorbing aerosols above Stratocumulus cloud decks in 2016-18. The algorithm will use a trained NN scheme for concurrent cloud and aerosol microphysical property retrievals that will be input to optimal estimation method. We will discuss the overall retrieval scheme, focusing on the input variables. Specifically, we use principle component analysis (PCA) to examine the information content available to describe the simulated cloud scenes (with adequate noise representation) by using various combinations of the measured and calculated polarimetric variables (total and polarized reflectance or a combination with degree of linear polarization). This sensitivity analysis will examine universal properties relevant to optimal retrieval of cloud properties using multiangle polarimetry.