Evaluation and comparison of JPSS VIIRS neural network retrievals of harmful algal blooms with other retrieval algorithms, validated against in-situ radiometric and sample measurements in the West Florida Shelf, and examination of impacts of atmospheric corrections, temporal variations and complex in-shore waters

We examine the potential for ocean color (OC) retrievals using a neural network (NN) technique recently developed by us to make up for the lack of a 678 nm florescence band on VIIRS, previously available on MODIS and important for Karenia brevis harmful algal bloom (KB HABs) retrievals.. NN uses VIIRS Remote Sensing Reflectance (Rrs) at 486, 551 and 671 nm to retrieve phytoplankton absorption at 443nm, from which both KB HABs and chlorophyll [Chla] concentrations can be inferred. NN retrievals are compared with retrievals obtained using other algorithms, including OCl/OCx and Semi-analytical algorithm for both complex and open ocean waters. VIIRS KB HABs retrievals in the WFS, using NN and other algorithms, are first compared against all co-incident in-situ cell count measurements available between 2012-16. Next, we compared retrievals obtained for different algorithms using in-situ radiometric Rrs measurements against sample measurements, 2017-18, for both the WFS and Atlantic coasts. Retrieval statistics showed (i) the important impact of short term (15-20 minutes) temporal variations and sample depth considerations in complex bloom waters. These limit satellite retrieval accuracies and utility; and (ii) particularly for high chlorophyll bloom waters, better retrieval accuracies were obtained with NN followed by OCI/OCx algorithms. Likely rationales: the longer Rrs wavelengths used with NN are less vulnerable (i) to atmospheric correction inadequacies than the deeper blue wavelengths used with other algorithms, and (ii) to spectral interference by CDOM in more complex waters.