Advancing Data and Tools for Machine Learning Modeling of Lake Water Quality

Abstract

Water quality is critical to lake ecosystem and human health. Increasingly, long term water quality data is becoming available which enables new modes of understanding water quality dynamics. Limnology has, as a field, entered an era where data-intensive modes of modeling, such as machine learning, have become viable and have the potential to offer novel insights into ecosystem dynamics. Here, I use time series modeling of dissolved oxygen in the surface waters of Lake Mendota (Wisconsin, USA) as a medium by which to discuss the state of our understanding of lake water quality drivers and propose novel modeling methods as pathways to knowledge acquisition. In Chapter 1, we introduce a benchmark dataset for lake water quality, LakeBeD-US. This dataset compiles data from multiple long-term monitoring institutions for distribu- tion in two formats: one conducive to ecological analyses and one for computer science applications. It is the first of its kind in the field of limnology to harmonize manually sampled and high-frequency data with the intent for use as a testing ground for novel machine learning methods. To showcase the applicability of this dataset for water qual- ity modeling with machine learning, we use a long short-term memory recurrent neural network to generate predictions of dissolved oxygen in the surface waters of Lake Men- dota. The machine learning model’s performance is comparable to existing process-based models. This dataset has great potential to be used in more complex ecological modeling tasks. In Chapter 2, we analyze the timescales at which drivers of dissolved oxygen concen- tration in Lake Mendota operate, and test a modular process-based model’s ability to recre- ate these multi-timescale dynamics. Interacting drivers operating at multiple timescales often pose a challenge for process-based model development, but our modular model per- forms well at recreating multiple scales of variability. The model does not recreate the dynamics of sub-daily timescales as well as longer timescales, which is informative as to how we can improve future iterations of the model. Our model is also designed to act as a modular compositional learning framework, which we propose as a potential method for addressing multiple scales of variability in a single water quality model.