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

Abstract

Water availability and sanitation are among the UN Sustainable Development goals for 2030. Remote sensing techniques are used to monitor and retrieve quality estimators from water bodies. Clean water is a scarce resource fundamental for human development and well-being. Treatment plants depend on the current water quality state to properly provide clean water. Combining laboratory measurements, provided by a water plant in Resistencia city, Argentina, and remote sensing data, i.e., surface reflectance, from Sentinel-2 platform, several algorithms were developed, trained, and compared for turbidity estimation. The model with the highest performance metrics was a random forest model, with Pearson’s coefficient of determination (R2) 0.918 and root-mean squared error (RMSE) 138.8 nephelometric turbidity units (NTU). Global feature importance and partial dependencies profiles techniques were applied to the random forest model to understand the spectral bands effects. Turbidity maps and time series were made and analyzed.