Measured records of suspended sediment concentrations are vital information to support water management activities. However, these measured time series are often incomplete and, as such, are not suitable for some analyses. This paper sets out the options for modelling suspended sediment concentrations to determine them in periods when measurements were not performed. The Danube River profile in Bratislava was selected as the case study. Regression using least absolute selection and shrinkage operator, support vector regression and deep learning neural network are compared in the paper to solve this task using various data sources. The results obtained show a significant increase in the precision of modelling suspended sediment concentrations over the standard method, which is a rating curve. Several variables were used to establish the suspended sediment concentration, because the same data as in this study may not be available everywhere. In particular, the use of climatic (precipitation and temperature) and hydrological inputs (flows) is assessed in order to promote the more general benefit of work. In the article, the authors propose an original method of modification of input climate data, which significantly increases the accuracy of modelling. The authors demonstrate that when using proposed methodology, the use of climate data, which are usually better available than hydrological data, resulted in a comparable degree of precision to standard modelling based on river flow data.
The aim is imputation of missing data.
Original method based solely on meteorological data is proposed.
Gridded meteorological databases, which have not yet been tested for this purpose, are used.
Machine learning is enhanced by so-called feature engineering.
Using solely meteorological data in modelling suspended sediment concentration has comparable accuracy as when river flows are used.