Evaluating Three Machine Learning Methods for Simulating Chlorophyll-a in Data-Scarce Lakes of Leon County, Florida

Chlorophyll-a (Chl-a), the primary photosynthetic pigment in algae, is widely used as a proxy for phytoplankton biomass and trophic status in lakes. While machine learning (ML) methods have been used for simulating Chl-a concentrations, there have been few studies examining ML methods for lakes with scarce monitoring data of water quality. This study evaluated simulation performance of three commonly used ML methods (i.e., Random Forest, XGBoost, and a Long Short-Term Memory (LSTM) network) for 11 data-scarce lakes in Leon County, FL. Each lake has only 1 – 6 sampling locations with a quarterly monitoring schedule, and there are a total of 1,012 sets of observations (43 to 252 per lake) for the study period of 2014 – 2024. Each observation set includes 43 water quality variables, e.g., phosphorus, organic nitrogen, and dissolved oxygen. Evaluating the three ML methods was based on three cross-validation operations and three statistical metrics (i.e., root mean square error, mean absolute error, and coefficient of determination). In a five-fold cross-validation, XGBoost achieved the best simulation performance. In a chronological cross-validation, all the methods had a similar modest simulation performance. A spatial cross-validation revealed that it was challenging to simulate Chl-a concentrations for two lakes with high Chl-a concentrations and large temporal variations. A method-agnostic importance test consistently identified a subset of water quality variables as top predictors across all the methods, including phosphorus, organic nitrogen, and biochemical oxygen demand. These findings demonstrate that ML is an effective tool for simulating Chl-a concentrations even for data-scarce lakes, whereas more monitoring data are needed for simulating high Chl-a concentrations with large temporal variations.