Machine learning methods integrating climate and water monitoring data to support modelling future water quality in lakes and reservoirs over decades

The dissertation studies the impact of climate change on water quality of freshwater ecosystems, with a particular focus on total phosphorus and water temperature. The PhD programme is part of the inventWater project, an Innovative Training Network (ITN) under the Marie Skłodowska-Curie Action (grant agreement No. 956623), dedicated to developing robust forecasting tools for water quality in rivers, lakes, and reservoirs at regional to global scales.

A distinct methodology is employed in three chapters of the dissertation to conduct the study. First, the relationship between extreme rainfall events and total phosphorus (TP) concentrations were examined in two rivers in Germany, namely Erft and Moehne. A combination of Generalized Additive Models (GAMs) and concentration-discharge (C-Q) relationships are used to analyze 15 years of daily TP data for the Erft River and 20 years of monthly TP data for the Moehne River. The outcomes reveal that TP concentrations consistently exceed expected values from C-Q relationships during heavy rainfall events and rainfall after dry periods. This suggests that intense rainfall significantly mobilizes nutrients. The study highlights the importance of considering the timing and intensity of rainfall in ecological conservation and water management. This study makes some innovative contributions to the field, including the analysis of both high-frequency and low-frequency TP data, the use of nonlinear regression methods, and the comprehensive assessment of TP responses to various rainfall patterns.

In a second study, future trajectories of TP loads to the world's 100 largest lakes were investigated under various shared socio-economic pathways (SSPs) and representative concentration pathways (RCPs). This study uses the WorldQual model to project TP loads from different sectors (domestic, manufacturing, agriculture, urban runoff, and background sources) until 2100 under three scenario combinations: SSP1-RCP2.6, SSP2-RCP6.0, and SSP5-RCP8.5. The results show a projected increase in global median TP load per river basin and lake area under SSP2-RCP6.0 and SSP5-RCP8.5, while SSP1-RCP2.6 suggests a declining trend after 2040. This underscores the influence of socio-economic development pathways on future nutrient pollution levels. The study further analyzes sectoral trends and regional variations in TP loading, identifying highly impacted and minimally impacted lakes and emphasizing the importance of targeted management strategies.

In a third study, the performance of machine learning (ML) models versus a process-based model was explored in predicting thermal dynamics in the Rappbode Reservoir, Germany. The chapter utilizes an extensive dataset of 9 million water temperature measurements collected from 2013 to 2022. Three ML models (Random Forest, XGBoost, and Long Short-Term Memory) are compared with a process-based model in terms of their ability to reproduce historical temperature profiles and project future thermal dynamics under the RCP8.5 climate scenario. The results demonstrate that the ML models effectively capture the seasonal temperature patterns and stratification dynamics of the reservoir, achieving accuracy comparable to the process-based model. However, the projections by the ML models are relatively insensitive to future warming compared to those by CE-QUAL-W2, suggesting a risk when extrapolating ML models beyond their training data range. This research highlights the potential and limitations of ML in understanding the impact of climate change on aquatic ecosystems, particularly in deep waters.

By integrating these three studies, this PhD thesis provides a comprehensive assessment of the multifaceted impacts of climate change, extreme rainfall events, and human activities on the water quality of rivers and lakes. The thesis emphasizes the need for sustainable management practices to mitigate the risks of eutrophication and ensure the long-term health of aquatic ecosystems. Furthermore, it showcases the strengths and limitations of different modelling approaches, including process-based models, statistical models, and machine learning, in understanding and projecting future water quality.