Neural net modeling of estuarine indicators: Hindcasting phytoplankton biomass and net ecosystem production in the Neuse (North Carolina) and Trout (Florida) Rivers, USA

Abstract

Phytoplankton biomass, as chlorophyll (Chl) a, and net ecosystem production (NEP), were modeled using artificial neural networks (ANNs). Chl a varied seasonally and along a saline gradient throughout the Neuse River (North Carolina). NEP was extremely dynamic in the Trout River (Florida), with phototrophic or heterotrophic conditions occurring over short-term intervals. Physical and chemical variables, arising from meteorological and hydrological conditions, created spatial and/or temporal gradients in both systems and served as interacting predictors for the trends/patterns of Chl a and NEP. ANNs outperformed comparable linear regression models and reliably modeled Chl a concentrations less than 20 m gL � 1 and NEP values, denoting the apparent non-linear interactions among abiotic and indicator variables. ANNs underestimated Chl a concentrations greater than 20 m gL � 1 , likely due to the periodicity of data acquisition not being sufficient to generalize system variability, the designated ‘lag’ effect for variables not being adequate to portray estuarine flow dynamics, the exclusion of (one or more) variables that would have improved prediction, and/or an unrealistic expectation of network performance. Variables indicative of meteorological and hydrological forcing and/or proxy measurements of phytoplankton had the greatest relative impact on prediction of Chl a and NEP. Except for their predictive capability, ANNs might appear to be of limited value for ecological applications and problem solving; interpreting the absolute impact of and/or interacting relationships among network