Inversion of water quality parameters in the South Bohemian Vltava River Basin based on machine learning algorithms (2000-2020)

Jonas Niese, BA

Master's thesis

Abstract:

Management of river water quality is important for mitigating eutrophication. After the water quality improvements since the 1990s, this trend might have turned in the upper Vltava catchment after 2010: Concentrations of the four optically inactive water quality parameters ammonia nitrogen, nitrate nitrogen, five-day biochemical oxygen demand, and total phosphorus might have increased, as results presented in this thesis suggest. All four parameters follow a land-use and ecosystem degradation gradient in the upper Vltava, increasing in concentration from the headwaters to the lower parts of the research area. In this study, five different types of algorithms (multiple stepwise linear regression, partial least squares regression, support vector regression, random forest regressor, and backpropagation artificial neural network) were performed on harmonized Landsat-Sentinel-2-data (HLS), in order to predict AN, NN, BOD5, and TP. Such machine learning methods can be useful tools for estimating the four named parameters, but they still bear a lot of challenges, especially given the often narrow streambed. Further improvements at all steps – from input data selection, over model optimization, to model selection – are required for ensuring accurate and robust predictions. The presented results indicate that this goal can be achieved. In most cases, machine learning algorithms clearly outperformed simpler linear models. Especially the method of the random forest regressor can often estimate a big fraction of variance, while also producing comparatively low errors. Of all four water quality parameters, NN was most effectively-predicted, with R2 = 0.555 by a random forest regressor, whereas AN is the most challenging with a maximum R2 = 0.156, also by a random forest regressor. BOD5 and TP prediction also remains challenging, but some models returned relatively strong metrics, indicating great potential for sound predictions of these parameters as well.

Keywords

Machine Learning, Water Quality, Eutrophication, Remote Sensing, Vltava, Neural Network, Random Forest, Landsat, Sentinel, Nitrogen, Phosphorus

Language used: English

Date on which the thesis was submitted / produced: 29. 3. 2024

Thesis defence

• Date of defence: 20. 5. 2024
• Supervisor: doc. Zhongbing Chen
• Reader: Jifeng Li

Czech University of Life Sciences Prague

Faculty of Environmental Sciences

Master programme:
Environmental Geosciences