Development of neural network retrievals of liquid cloud properties from multi-angle polarimetric observations

We present a neural network (NN) based algorithm for the retrieval of liquid low-level marine stratocumulus cloud microphysical property parameters (cloud optical depth, cloud droplet size effective radius and variance) from airborne multi-angle polarimetric measurements. We establish our retrieval method for the Research Scanning Polarimeter (RSP) airborne instrument, which measures both polarized and total reflectance in the spectral range of 410–2260 nm, scanning along the flight track at ∼150 viewing zenith angles spanning the angular range between −60° and 60°. In this study, we present the development of the algorithm, including the optimization and selection of input parameters and the network architecture. We perform a sensitivity study to test the effect of random and correlated instrument noise on the retrieval performance, and to assess which of the measured radiometric quantities (i.e., total reflectance, polarized reflectance, degree of linear polarization and combinations thereof) are best suited for marine stratocumulus liquid cloud property retrievals using simulated RSP data. Finally, we show the application of this method to airborne observations from the ObseRvations of Aerosols above CLouds and their intEractionS (ORACLES) 2016 field campaign, which primarily encountered low altitude marine clouds. Retrieved cloud optical depth compares favorably (r2 = 0.96) to standard algorithms, but cloud droplet size effective radius less so (r2 = 0.45), providing an assessment of the NN approach strengths and limitations. Specifically, the latter seemed to be affected by the cloud macro-structure and the liquid cloud droplet vertical distribution.