Estimating the Model Parameters for Remote Sensing Reflectance Pixel by Pixel: a Neural Network Approach for Optically Deep Waters

Remote sensing reflectance (R (rs)) and inherent optical properties (IOPs) conversions are fundamental in accurate satellite measurements to guarantee the semi-analytical retrieval quality of IOPs and the biogeochemical products. Traditionally, the R (rs)-IOPs conversions are determined by a quadratic polynomial function with two model parameters (G (x = 0,1)). However, G (x) values vary in time and location, which are attributed to the spatial and temporal variability inherent to illumination conditions, sea surface properties, and meteorological states. To improve the performance of three classical existing models used for R (rs)-IOPs conversions, we designed two novel neural network models (NNG (x = 1,2)) to quantitatively calculate G (x) from the R (rs) spectrum pixel by pixel without requirement of any auxiliary illumination and meteorological data, and then proposed for R (rs)-IOPs conversions. We evaluated these approaches with numerical simulations and field measurements, and the results show that the NNG (x) models are more effective in semi-analytically converting R (rs) into IOPs than the three existing models. Furthermore, we applied the NNG (x) models to satellite images to understand the downstream influence of the G (x) values on IOPs estimates for the global oceans. We further confirm that the G (x) values dramatically change for the global ocean, which is especially true for very oligotrophic gyres, coastal waters, and high latitude oceans. When we use a constant G (x) for the R (rs)-IOPs conversions, it leads to substantial uncertainty of up to 30% in the IOPs retrievals for China's coastal regions. Our results suggest that it is possible to improve the data quality of IOPs for the global oceans by providing accurate pixel-level G (x) values using NNG (x) models.