Estimation of Absorption Parameters for Coastal Waters Using Hyperspectral Imaging

This study investigates the synergies between machine learning and deep learning techniques to enhance the estimation of water column properties Total Suspended Matter (TSM) and Colored Dissolved Organic Matter (CDOM), and location-specific spectral slopes of non-algal particles (s
nap
) and CDOM (s
cdom
). These water parameters are vital for characterizing water composition, monitoring water quality, understanding optical properties, and contributing to broader environmental health assessments. We propose a probabilistic model based on a well-known radiative transfer model for reflectance data to estimate these critical parameters. Our experimental analysis on simulated hyperspectral data assesses the suitability of this model compared to Partial Least Squares Regression(PLSR) and Convolutional Neural Network (CNN) for accurate estimation of both TSM and s
nap
, as well as CDOM and s
cdom
for single spectra. Additionally, we evaluate the likelihood-based model’s performance for multiple spectra. The proposed likelihood-based model shows the lowest prediction errors for s
nap
and s
cdom
. Moreover, CNN-based regression does reasonably well in estimating water constituents TSM and CDOM.