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

Chlorophyll-a (Chl-a) is widely used as a proxy for phytoplankton biomass and lake trophic status. This exploratory study compared Random Forest, XGBoost, and a one-step Long Short-Term Memory (LSTM) network using 1012 station–date observations and 43 contemporaneous water quality predictors from Leon County, Florida, during 2014–2024. Initial shuffled five-fold validation favored XGBoost (mean RMSE 10.45 µg L−1; mean R2 0.484), but shared stations across folds and prior manual configuration selection limit that comparison. Additional validation held out entire stations in five outer folds and selected configurations using three station-blocked inner folds. Pooled outer RMSE was 11.92, 10.49, and 10.91 µg L−1 for Random Forest, XGBoost, and LSTM, respectively; the corresponding R2 values were 0.371, 0.513, and 0.473. Six expanding training periods ending in 2017–2022, each assessed over the next two years, evaluated sensitivity to chronological cutoff. These tests complement the original 2014–2019 versus 2020–2024 split, for which R2 was 0.054–0.071. Random Forest had the lowest RMSE in all six windows, while LSTM had negative R2 in four. Spatial transfer remained limited for Lakes Munson and Iamonia. Nested evaluation separates the new configuration search from each outer assessment, but the dataset and predictor panel had already been examined; the reanalysis is not an untouched external validation. The findings describe conditional concentration estimation using contemporaneous measurements, not advance forecasting or assured transfer across contrasting lakes.