Global distribution of the spectral power coefficient of particulate backscattering coefficient obtained by a neural network scheme

Abstract

The spectral power coefficient (η) of the particulate backscattering coefficient (bbp(λ)) is directly estimated from the remote sensing reflectance (Rrs(λ)) with a neural network-based scheme (NNη) in this study. Evaluations with both synthetic dataset and in-situ measurements show that NNη could significantly improve the accuracy of estimated η compared to several conventional schemes reported in the literature that are based on chlorophyll-a concentration (Chl), band ratios of Rrs(λ), or remotely sensed bbp(λ). Demonstrations with measurements from MODerate resolution Imaging Spectroradiometer (MODIS) Aqua further confirm the robustness of NNη, where reasonable spatial distribution and seasonality of η in the global oceans can be acquired by NNη. High and low η values are observed in the oligotrophic gyres and the coastal zones, respectively, which are consistent with the current understanding of η distribution concluded from theoretic analysis and repeated field measurements in the global ocean. Implementation of NNη to 19-year MODIS monthly composite measurements from 2003 to 2021 reveals strong seasonal variations of η in most of the global ocean, but the decadal changes of η are insignificant in the majority (∼82.2%) of the global ocean. Similar to any empirical algorithms, the performance of NNη is dependent on the training dataset, particularly its range, here a proper upper limit of η for natural waters is provided.