Hidden Signals of Eutrophication: Artificial Intelligence and the Phytoplankton–Organic Matter Connection

Anthropogenic inputs of nutrients and organic matter can promote coastal eutrophication, phytoplankton blooms, and oxygen depletion through nonlinear environmental interactions. Araruama Lagoon (Rio de Janeiro State, Brazil) is a perennial hypersaline lagoon with a mean salinity of 52 PSU and limited water exchange. It has experienced blooms associated with surrounding urbanization.This study evaluated artificial neural networks for predicting diatom, dinoflagellate, and cyanobacteria abundance from salinity, dissolved oxygen, temperature, pH, turbidity, and apparent color. The initial training datasets comprised 438, 437, and 428 observations, respectively. The corresponding validation datasets contained 45, 43, and 43 observations. MinMax scaling was fitted to the training data and applied to the validation data. EllipticEnvelope was fitted separately to both datasets using the six predictors and a contamination parameter of 0.10. Flagged observations were classified as statistically atypical, without being considered confirmed measurement errors. TensorFlow/Keras multilayer perceptrons comprised two hidden layers of 32 neurons each, ReLU activation, 20% dropout, and a linear output layer. Minimum validation loss determined weight selection, and performance was assessed using the same validation datasets, without a separate test dataset. Before outlier removal, validation R² values were 0.834, 0.886, and 0.580 for diatoms, dinoflagellates, and cyanobacteria, respectively, already indicating promising predictive performance, particularly for diatoms and dinoflagellates. After removing flagged observations from both datasets and retraining the models, R² values reached 0.902, 0.929, and 0.630. Because the composition of the validation datasets also changed, these results are presented as an exploratory sensitivity check rather than as directly comparable performance estimates. The study contributes to environmental geochemistry by relating phytoplankton abundance to conditions associated with eutrophication and organic matter cycling. Phytoplankton is an important source of autochthonous organic matter, whose decomposition influences oxygen availability and nutrient dynamics. Although the models do not incorporate direct organic geochemical measurements, predicting phytoplankton abundance helps characterize an important biological source of autochthonous organic matter and supports the interpretation of its production and transformation in eutrophic lagoons. Further work should incorporate organic geochemical measurements, nutrient and hydrodynamic variables, and assess model generalizability using independent datasets not involved in model selection.