Neural network approach for correction of multiple scattering errors in the LISST-VSF instrument

Abstract

The LISST-VSF is a commercially developed instrument used to measure the volume scattering function (VSF) and attenuation coefficient in natural waters, which are important for remote sensing, environmental monitoring and underwater optical wireless communication. While the instrument has been shown to work well at relatively low particle concentration, previous studies have shown that the VSF obtained from the LISST-VSF instrument is heavily influenced by multiple scattering in turbid waters. High particle concentrations result in errors in the measured VSF, as well as the derived properties, such as the scattering coefficient and phase function, limiting the range at which the instrument can be used reliably. Here, we present a feedforward neural network approach for correcting this error, using only the measured VSF as input. The neural network is trained with a large dataset generated using Monte Carlo simulations of the LISST-VSF with scattering coefficients 𝑏=0.05−50m−1, and tested on VSFs from measurements with natural water samples. The results show that the neural network estimated VSF is very similar to the expected VSF without multiple scattering errors, both in angular shape and magnitude. One example showed that the error in the scattering coefficient was reduced from 103% to 5% for a benchtop measurement of natural water sample with expected 𝑏=10.6m−1. Hence, the neural network drastically reduces uncertainties in the VSF and derived properties resulting from measurements with the LISST-VSF in turbid waters.