Retrieving bbp and POC from CALIOP: A deep neural network approach

The CALIOP instrument has demonstrated tremendous potential in terms of oceanic detection. Nighttime measurements from spaceborne lidar increase knowledge of the diurnal variability of ocean optical properties, while measurements in polar regions fill the gaps that passive measurements of ocean color may leave in these areas. However, due to non-ideal instrumental characteristics, errors may occur during retrieval of the ocean optical properties that are derived directly from CALIOP's measured signals. These errors are caused by the transient response (TR) function of the detectors and the polarization crosstalk (CT) between the polarization beam splitters. However, there is no study that integrates and explicitly describes both the required corrections. In addition, the published physics-based methods for retrieving the particulate backscatter coefficient (bbp) from CALIOP use an empirical linear relationship between bbp and the backscatter coefficient at 180° b(π) without considering the spatial and temporal variability associated with the conversion coefficient. In this study, after both effects are eliminated through TR deconvolution and CT correction, an innovative deep neural network (DNN) approach is proposed to retrieve bbp and particulate organic carbon (POC) from CALIOP. The DNN is trained with the CALIOP total ocean column-integrated depolarization ratio and subsurface column-integrated perpendicular backscatter data, which are combined with the collocated daytime MODIS bbp and POC products. Once the DNN has been trained with the daytime MODIS product, measurements can be derived from CALIOP in conditions not covered by MODIS data, such as at night or in polar regions. Compared to published physics-based methods, this approach does not require diffuse attenuation coefficients, nor are additional assumptions regarding the conversion coefficient between bbp and the backscatter coefficient at 180° b(π) necessary. A comparison of global total depolarization ratios, which are calculated before and after TR deconvolution and CT correction, indicates the necessity and efficacy of correction, and the preliminary results show that the oceanic properties derived from CALIOP via this DNN-based approach agree better with the MODIS products than those derived with the traditional physics-based methods. The use of lidar measurements to derive bbp and POC without relying on passively acquired data is of great interest to the ocean color community.