Comparing the performance of 10 machine learning models in predicting Chlorophyll a in western Lake Erie

Area of Science:

• Environmental Science
• Data Science
• Limnology

Background:

• Algal blooms pose significant global threats, exacerbated by climate change and eutrophication.
• Machine learning models (MLMs) show promise for predicting algal blooms, but their direct simulation performance in large freshwater systems like the Great Lakes is understudied.

Purpose of the Study:

• To evaluate and compare the predictive performance of 10 popular MLMs for Chlorophyll a (Chl a) concentration in western Lake Erie.
• To identify optimal feature combinations and assess the impact of data preprocessing techniques like outlier removal on MLM accuracy.

Main Methods:

• Utilized 15 water quality parameters from 2012-2022 to train and test 10 MLMs for Chl a prediction.
• Performed exhaustive feature selection testing (32,767 combinations) and analyzed the impact of outlier removal.

Main Results:

• Outlier removal substantially improved prediction accuracy, increasing R² from 0.35 to 0.84 for the Gradient Boosting Decision Trees (GBDT) model.
• Feature selection enhanced MLM performance, with Polynomial Regression R² improving from 0.71 to 0.82.
• GBDT (R² = 0.84) and Random Forest (R² = 0.82) demonstrated superior performance in predicting Chl a.

Conclusions:

• This study establishes a benchmark for MLM performance in predicting Chl a in western Lake Erie.
• Particulate organic nitrogen (PON) was identified as the most critical predictor for Chl a concentration.
• Optimized feature selection and data preprocessing enhance the effectiveness of water quality monitoring for sustainable management.