VIIRS High Spatial Resolution Ocean Color Data Derived Using the Deep Convolutional Networks

The Visible Infrared Imaging Radiometer Suite (VIIRS), which is onboard both the Suomi National Polar-orbiting Partnership (SNPP) and National Oceanic and Atmospheric Administration (NOAA)-20 satellites, has 22 spectral bands acquired at two different spatial resolutions: 16 moderate resolution bands (M-bands) with a spatial resolution of 750 m at nadir, and 5 high spatial resolution imaging bands (I-bands) of 375 m at nadir. Since the wavelengths of the I-bands are close to some of those in the M-bands, the high spatial resolution structures of I-bands can be used to super-resolve the M-band imagery from 750 m to 375 m resolution. For example, VIIRS-derived normalized water-leaving radiance nL_w(λ) at the I1 band (wavelength of 638 nm) has strong signals in coastal regions, and it can be used to super-resolve VIIRS-derived nL_w(λ) spectra at M1-M6 bands with wavelengths from 410 to 745 nm. In this study, deep Convolution Neural Networks (CNN) is employed to perform the super-resolution work. Specifically, the CNN is trained with nL_w(l) images at I1 band in the Bohai Sea and Baltic Sea, and the models are used to super-resolve nL_w(λ) images at the M1-M6 bands. The sensitivity of super-resolution performance to the spectral wavelength from M1 to M6 in both regions is tested. The CNN model trained from the Bohai Sea is also applied to the Baltic Sea (and vice versa), so that the dependency of the algorithm performance on specific ocean region can be evaluated. It is demonstrated that new improved VIIRS high spatial resolution ocean color data provide more information of coastal water property and dynamics.