MODELING PHYTOPLANKTON ABUNDANCE IN SAGINAW BAY, LAKE HURON: USING ARTIFICIAL NEURAL NETWORKS TO DISCERN FUNCTIONAL INFLUENCE OF ENVIRONMENTAL VARIABLES AND RELEVANCE TO A GREAT LAKES OBSERVING SYSTEM1

Abstract

Abstract Phytoplankton abundance, as chl a, in Saginaw Bay, Lake Huron was modeled using artificial neural networks. Suites of abiotic variables served as predictors for the trends/patterns in chl a concentrations. Spatial and temporal gradients of sampling stations throughout the bay were evident, with physical/chemical differences arising from hydrological/meteorological forcing and zebra mussel recruitment. Chlorophyll a concentrations displayed corresponding disparities; concentrations differed between the inner and outer bays and varied intra‐ and inter‐annually. Trained networks reproduced the intrinsic variance and magnitude of chl a dynamics. Modeled‐measured concentrations best approximated a 1:1 relationship in a hybrid network incorporating both supervised and unsupervised training whereas concentrations greater than 15 μg/L were underestimated in networks utilizing only supervised training, likely because of inadequate training data. Variables indicative of phytoplankton nutrition, acting as proxy measurements of algal biomass, and/or corresponding to descriptors of hydrological and meteorological forcing had the greatest influence upon modeled concentrations. A conjunctive decision tree and a novel sensitivity analysis provided rule‐based information and comprehensible interpretation of relationships among multiple predictor variables. From this, the “knowledge” embedded in trained networks proved extractable and usable for ecological theory generation and/or decision making within water‐quality problem solving. Forecasting initiatives within the developing Great Lakes Observing System may be best served by embedding neural networks in mechanistic models to quantitatively initialize variables, qualitatively delineate conditions for projecting ecological structure, and/or estimate deviations from predictability within mechanistic simulations.