Application of machine learning techniques to derive sea water turbidity from Sentinel-2 imagery

Abstract

Earth Observation (EO) from satellites has the potential to provide comprehensive, rapid and inexpensive information about water bodies, integrating in situ measurements. Traditional methods to retrieve optically active water quality parameters from satellite data are based on semi-empirical models relying on few bands, which often revealed to be site and season specific. The use of machine learning (ML) for remotely sensed water quality estimation has spread in recent years thanks to the advances in algorithm development and computing power. These models allow to exploit the wealth of spectral information through more flexible relationships and are less affected by atmospheric and other background factors. The present study explores the use of Sentinel-2 MultiSpectral Instrument (MSI) Level-1C Top of Atmosphere spectral radiance to derive water turbidity, through application of machine learning techniques. A dataset of 222 combination of turbidity measurements, collected in the North Tyrrhenian Sea – Italy from 2015 to 2021, and values of the 13 spectral bands in the pixel corresponding to the sample location was used. Two regression techniques were tested and compared: a Stepwise Linear Regression (SLR) and a Polynomial Kernel Regression. The two models show accurate and similar performance (R2 = 0.736, RMSE = 2.03 NTU, MAE = 1.39 NTU for the SLR and R2 = 0.725, RMSE = 2.07 NTU, MAE = 1.40 NTU for the Kernel). A band importance analysis revealed the contribution of the different spectral bands and the main role of the red-edge range. The work shows that it is possible to reach a good accuracy in turbidity estimation from MSI TOA reflectance using ML models, fed by the whole spectrum of available bands, although the possible generation of errors related to atmospheric effect in turbidity estimates was not evaluated. Comparison between turbidity estimates obtained from the models with turbidity data from Copernicus CMEMS dataset named ‘Mediterranean Sea, Bio-Geo-Chemical, L3, daily observation’ produced consistent results. Finally, turbidity maps from satellite imagery were produced for the study area, showing the ability of the models to catch extreme events.