On an Artificial Neural Network Approach for Predicting Photosynthetically Active Radiation in the Water Column

Abstract

About 1,600 bio-geo-chemical Argo floats (BGC-Argo), equipped with a variety of physical sensors, are currently being deployed in the ocean around the world for profiling the water characteristics up to a depth of 2,000 m. One of the parameters measured by the Argo is the radiometric measurement of downward irradiance, which is important for primary production studies. The multispectral Ocean Color Radiometer measures the downwelling irradiance at three wavelengths 380 nm, 412 nm and 490 nm plus the photosynthetically available radiation (PAR) integrated from 400 nm to 700 nm. This study proposes a method to reconstruct the PAR sensor values from readings of the remaining onboard sensors, independent of the location the BGC-Argo is being deployed. This allows for the PAR channel being replaced by a fourth band in the visible range. Stahl et al. [1] have already shown, that a machine learning approach, based on a multiple linear regression (MLR) or on a regression tree (RT), is capable of predicting the PAR values based on other parameters measured by the physical sensors of the BGC-Argo float. In this study, a nonlinear Artificial Neural Network (ANN) was used for the prediction of PAR. The ANN achieved a better coefficient of determination R² of 0.9968, compared with the MLR approach, which achieved an R² of about 0.97 for a combined dataset consisting of measurements from three different geographical locations. Therefore, it was concluded that the ANN was better suited to generalise the underlying transfer function.