Estimating phytoplankton group abundances in an agricultural pond from in situ sensed data with machine learning: use of the SHAP analysis for ecological assessments

Phytoplankton are a crucial component of aquatic ecosystems and are closely tied to water quality. Direct counts of phytoplankton abundances are resource-demanding, but the indirect estimation of those abundances has proven to be beneficial when conducting ecological assessments of waterbodies. Agricultural ponds serve as important water sources for irrigation, recreation, processing harvested agricultural products, animal watering, and other purposes. This work examined the use of random forest (RF), coupled with a Shapley Additive exPlanations (SHAP) analysis, to estimate the abundances of phytoplankton groups in an agricultural pond in Maryland. In situ sensing (ISS) of water quality parameters on a permanent sampling grid during the produce growing season provided dissolved oxygen, pH, specific conductance, chlorophyll a, phycocyanin, fluorescent dissolved organic matter, and turbidity measurements. Phytoplankton abundance data was determined using a modified Utermöhl microscopy method. Values of the determination coefficient for training and testing datasets were on average 0.81 and 0.74, and varied from 0.50 to 0.88 for ISS predictors, respectively. The explanatory analysis using SHAP revealed that the most influential predictors, identified as the top three for each phytoplankton taxonomic group, were specific conductance, fluorescent dissolved organic matter, and chlorophyll a. The RF analysis provided good estimates of the abundance of the phytoplankton community in agricultural pond waters and the addition of the SHAP analysis allowed for an exploration of what factors were most critical in supporting the phytoplankton groups observed.