Turbidity estimation by machine learning modelling and remote sensing techniques applied to a water treatment plant

Abstract

Clean water is a scarce resource, fundamental for human development and well-being. Remote sensing techniques are used to monitor and retrieve quality estimators from water bodies. In situ sampling is an essential and labour-intensive task with high costs. As an alternative, a large water quality dataset from a potabilisation plant can be beneficial to this step. Combining laboratory measurements from a water treatment plant in North-East Argentina and spectral data from the Sentinel-2 satellite platform, several regression algorithms were proposed, trained, and compared for turbidity estimation at the plant inlet water in a local river. The highest performance metrics were from a Random Forest model with a coefficient of determination close to 1 (0.913) and the lowest root-mean-squared error (143.9 nephelometric turbidity units). Global feature importance and partial dependencies profile techniques identified the most influential spectral bands. Maps and histograms were made to explore the spatial distribution of turbidity.