Sentinel-2 imagery coupled with machine learning to modelling water turbidity in the Doce River Basin, Brazil

Abstract

Remote sensing and machine learning are techniques that can be used to monitor water quality properties, surpassing the limitations of the conventional techniques. Turbidity is an important water quality property directly influenced by the Fundão dam tailing rupture, which spilled tons of ore tailing in rivers of the Doce River Basin, Southeastern Brazil. We tested different machine learning algorithms coupled with 10 m resolution Sentinel-2 images, to model and spatially predict the water turbidity of the Doce basin rivers affected by the Fundão dam rupture. Results indicate that the cubist model presented the best performance. Both single bands and spectral indices presented great importance for modelling water turbidity. In particular, the Fe3 index (simple ratio between red and blue bands) was the most important covariate, highlighting the spectral response of the suspended sediments rich in Fe oxides. The red and near-infrared bands were the most relevant single bands for modelling turbidity, since the great spectral contrast between clean and turbid water in these bands. The water turbidity was considerably higher in the rainy season and for the upstream Doce basin where the Gualaxo do Norte and Carmo rivers are located. This is associated with the great deposition of the Fundão dam tailings inside and outside these rivers, besides the hydraulic and geomorphological characteristics of the Gualaxo do Norte and Carmo sub-basins.