Estuary water quality classification through deep learning image segmentation, an example of Hangzhou Bay

Suspended sediment plays a dual role in the aquatic ecosystems. It serves an important part of the aquatic nutrient cycle but is also a pollutant in aquatic ecosystems. Traditional methods of measuring suspended sediment concentration (SSC) are mainly in-situ sampling and laboratory analysis. Through optical sensors carried by satellites or airplanes, large-scale SSC is possible to be acquired, which helps to improve our understanding of connections of source, sink and pathway. Although lots of efforts have been done in water quality modeling, methods utilized are mainly multi-layer perceptrons, machine learning algorithms or physio-chemical models, few tries on convolutional neural networks have been done. Here we take the Landsat image of Hangzhou Bay as an example. Based on the inversion algorithm the water quality classification is conducted through by 1D-CNN and 2D-UNet model. Best match between the band and reflectance of SSC received by the sensor are carefully selected as input. Both models prove feasible with convincing accuracy and precision, despite some discrepancies in details. Water quality classification could be furtherly completed if taking multi-index of water into account.