A Neural Network Model for K(λ) Retrieval and Application to Global Kpar Monitoring

Abstract

Accurate estimation of diffuse attenuation coefficients in the visible wavelengths K d(λ) from remotely sensed data is particularly challenging in global oceanic and coastal waters. The objectives of the present study are to evaluate the applicability of a semi-analytical K d(λ) retrieval model (SAKM) and Jamet’s neural network model (JNNM), and then develop a new neural network K d(λ) retrieval model (NNKM). Based on the comparison of K d(λ) predicted by these models with in situ measurements taken from the global oceanic and coastal waters, all of the NNKM, SAKM, and JNNM models work well in K d(λ) retrievals, but the NNKM model works more stable and accurate than both SAKM and JNNM models. The near-infrared band-based and shortwave infrared band-based combined model is used to remove the atmospheric effects on MODIS data. The K d(λ) data was determined from the atmospheric corrected MODIS data using the NNKM, JNNM, and SAKM models. The results show that the NNKM model produces <30% uncertainty in deriving K d(λ) from global oceanic and coastal waters, which is 4.88-17.18% more accurate than SAKM and JNNM models. Furthermore, we employ an empirical approach to calculate K par from the NNKM model-derived diffuse attenuation coefficient at visible bands (443, 488, 555, and 667 nm). The results show that our model presents a satisfactory performance in deriving K par from the global oceanic and coastal waters with 20.2% uncertainty. The K par are quantified from MODIS data atmospheric correction using our model. Comparing with field measurements, our model produces ~31.0% uncertainty in deriving K par from Bohai Sea. Finally, the applicability of our model for general oceanographic studies is briefly illuminated by applying it to climatological monthly mean remote sensing reflectance for time ranging from July, 2002- July 2014 at the global scale. The results indicate that the high K d(λ) and K par values are usually found around the coastal zones in the high latitude regions, while low K d(λ) and K par values are usually found in the open oceans around the low-latitude regions. These results could improve our knowledge about the light field under waters at either the global or basin scales, and be potentially used into general circulation models to estimate the heat flux between atmosphere and ocean.

A Spectrally Selective Attenuation Mechanism‐Based Kpar Algorithm for Biomass Heating Effect Simulation in the Open Ocean

“…To overcome this problem, the Hydrolight model was used to generate two synthetic data sets with SeaWiFS wavelengths in this study. First, a synthetic data set was used to refine the four‐band neural network K d (λ) algorithm (FNNK) (Chen et al, ) to improve the capability of FNNK in reducing the R rs residual error in the satellite data. This data set has 12,021 data points (see Table ), with chlorophyll‐a randomly changing from 0.01 to 15 mg m −3 , large suspended particles randomly ranging from 0.01 to 10 g m −3 , the gelbstoff absorption coefficient at 443 nm randomly ranging from 0.001 to 1.0 m −1 , both exponential slopes ( S ) associated with gelbstoff and detritus randomly fixed in the range of 0.01–0.02, and power coefficient of backscattering coefficient ( Y ) randomly fixed in the range of 0–2.…”

“…Chen et al () showed that the influences of R rs residual errors could be removed as long as the spectral characteristics of R rs residual errors are properly assimilated into the ocean color algorithm. In this study, the FNNK algorithm originally developed by Chen et al () is refined to obtain the vertical diffuse attenuation coefficient ( K d (λ)). Instead of single band method in the original FNNK algorithm (SFNNK) (Chen et al, ), the band difference method proposed by Chen et al () is coupled into the FNNK algorithm to minimize the impacts of R rs residual errors on K d (λ) estimations.…”

“…In this study, the FNNK algorithm originally developed by Chen et al () is refined to obtain the vertical diffuse attenuation coefficient ( K d (λ)). Instead of single band method in the original FNNK algorithm (SFNNK) (Chen et al, ), the band difference method proposed by Chen et al () is coupled into the FNNK algorithm to minimize the impacts of R rs residual errors on K d (λ) estimations. Further, the BFNNK algorithm is initialized using residual error‐included R rs data, so it can reduce some influences of the R rs residual errors and achieve more accurate results than the SFNNK algorithm.…”

“…To accurately predict the impacts of biological and physical processes on the vertical changes in solar radiation and bioheating effects in the upper ocean (Mobley & Boss, ; Oschlies, ; Zhang et al, ), the vertical diffuse attenuation coefficient of PAR ( K par ) should first be determined. The value of K par was commonly presumed to have no relationship with depth ( z ) in many studies (Buiteveld, ; Chen et al, ; Saulquin et al, ; Son & Wang, ; Zheng et al, ); that is to say, a constant is used that is equal to the mean K par value within a given “depth.” However, it has been previously demonstrated that even for pure water, the K par value varies markedly with depth (Lee, ; Mobley, ; Mobley & Boss, ; Wei & Lee, ). Following the radiance transfer simulation provided by Hydrolight version 5.2 (Mobley & Sundman, ), Figure shows that K par decreases with depth from 0 to 50 m, with the maximum value at the surface.…”

Spatial pattern of K&lt;sub&gt;d&lt;/sub&gt;(PAR) and its relationship with light absorption of optically active components in inland waters across China

“…However, in situ measurements of Kd(PAR) in waters have obvious limitations, and it is difficult to achieve spatial coverage. Satellite remote sensing has achieved the mapping of Kd(PAR) distribution from various types of satellite remote sensing data in open sea, coastal and inland waters in recent years (Chen & Zhu & Wu & Cui & Ishizaka & Ju, 2015;Shi et al, 2014;Song et al, 2017). However, Environmental change and anthropogenic activity have made it challenging to accurately assess Kd patterns in the extremely turbid inland waters (Zheng & Ren & Li & Huang & Liu & Du & Lyu, 2016).…”