Proceedings Article | 18 October 2019
KEYWORDS: Satellites, In situ metrology, Neural networks, Atmospheric corrections, Optical simulations, Optical properties, Visible radiation, Infrared imaging, Radiometry, Remote sensing
Our previous work showed the efficacy neural network (NN) approaches for satellite detection of Karenia brevis (KB) harmful algal blooms (HABs) in the West Florida Shelf (WFS). Applying a multiband NN, trained on a wide range of synthetically simulated inherent optical properties (IOPs), the NN takes Visible Infrared Imaging Radiometer Suite (VIIRS) remote sensing reflectance (Rrs) measurements at 486, 551 and 671 nm bands as inputs and produces related IOPs at 443 nm, including the absorption coefficient of phytoplankton, (aph443) as retrieved output images. We convert the retrieved aph443 values into equivalent chlorophyll-a [Chla] and KB HAB concentration mages, using known empirical relationships. We then compared VIIRS NN retrievals of KB HABs in the WFS with those obtained using other algorithms, and evaluated their accuracy against near co-incident in situ cell count measurements available over the 2012-16 period. Results highlighted the important impact of rapid (15-20 minutes) temporal variations on retrieval accuracy, and showed that the NN technique exhibited the highest accuracy statistics for retrievals of KB HABs in the WFS. This paper presents significant results in three areas.
(i) We extend evaluations and comparisons of NN retrievals of aph443 and [Chla] with retrievals from other ocean color (OC) algorithms to waters beyond the WFS, including both complex coastal and open ocean waters, along Florida and Atlantic coasts, with a large dynamic range of chlorophyll-a values. We also now add in situ radiometric measurements to obtain Rrs inputs to retrieval algorithms. In contrast to satellite Rrs retrievals, these are invulnerable to atmospheric transmission correction errors. This permits comparison of in situ radiometric measurement-based retrievals with simultaneous co-located satellite retrievals and validation against in situ sample measurements. This allows us to isolate different factors affecting retrieval accuracy and evaluate the intrinsic merits of different algorithms unencumbered by inadequate/erroneous atmospheric transmission assumptions and/or satellite instrumental calibration limitations. Results obtained extend and demonstrate the efficacy of NN algorithms to widely varying waters beyond the WFS,
(ii) Since it is conjectured that in satellite retrievals, it is the deep blue wavelengths that are more detrimentally affected by atmospheric correction inadequacies, we examined impacts on algorithm retrieval accuracy when deep blue wavelengths are used for retrieving Rrs values. Retrievals using NN and OCI/OCx algorithms were compared against in situ sample measurements, using in situ radiometric Rrs measurements and satellite Rrs retrievals as inputs, first at 443 nm (deep blue) and then at 486 nm (non-deep blue). The results unambiguously show that satellite retrieval accuracy, as well as intrinsic retrieval accuracy from in situ radiometric measurements are improved when deep blue wavelengths (443 nm) are avoided in favor of non-deep blue wavelengths (486 nm), raising issues of both atmospheric correction as well as possible underwater spectral interference in CDOM rich and complex waters. Thereby arguing for use of OC algorithms using the longer wavelengths.
(iii) Finally, new quantitative analysis of temporal, intra pixel and sample depth variability highlights their important impact on retrieval accuracy.
