Effects of meteorological forcing on coastal eutrophication: Modeling with model trees Androniki Tamvakis * , John Miritzis, George Tsirtsis, Alexandra Spyropoulou, Sofie Spatharis University of the Aegean, Department of Marine Sciences, University Hill, 81100 Mytilene, Greece article info Article history: Received 26 May 2012 Accepted 20 September 2012 Available online 4 October 2012 Keywords: process-based modeling linear regression machine learning algorithm model tree dynamics eutrophication abstract The exploration of processes leading to coastal eutrophication is a major challenge in ecological research, particularly in light of important new policies such as the European Water Framework Directive. In the present study primary production (in terms of chlorophyll a e chl a) is modeled based on a number of abiotic parameters using model trees (MTs), a machine learning (ML) approach whereby linear regres- sions are induced within homogeneous subsets of samples (tree leaves). Standardized regression was applied to determine the relative weight of abiotic parameters in the MT tree leaves whereas the effi- ciency of the MT method in chl a prediction was tested against neural networks (NNs) which is the most frequently used ML approach, and the classical multiple linear regression (MLR). To assess the efficiency of models to describe eutrophication-related responses under different environmental conditions, the methods were applied on a coastal ecosystem affected by terrestrial runoff for two meteorologically contrasting annual cycles: a typical dry (’04e’05) and a typical wet (’09e’10). MTs showed increased predictive power in chl a prediction attributed to the discrimination of input data space into tree leaves, instead of using a uniform space as in NNs and MLR. By grouping samples of each tested annual cycle (wet and dry) on a seasonal basis into discrete groups/leaves, MTs offer a much more explanatory description of ecosystem status than NNs and MLR. The discriminating variables forming tree leaves and the weighing coefficients of Linear Models (LMs) in each leaf provided a useful scaling of abiotic parameters driving chl a dynamics. The MT method is thus proposed as an efficient tool for obtaining insights into ecosystem processes leading to eutrophication events in coastal ecosystems and a useful component in integrated coastal zone management. Ó 2012 Elsevier Ltd. All rights reserved. 1. Introduction Coastal areas worldwide are increasingly susceptible to eutro- phication phenomena often due to anthropogenic causes such as sewage and terrestrial runoff (Beman et al., 2005). Recently, coastal eutrophication has received special attention in light of new poli- cies e.g. the Water Framework Directive 2000/60/EC, the protocol for integrated coastal zone management and marine biodiversity protection (Karydis, 1996; Coll et al., 2010; Ruiz and Velasco, 2010). However, eutrophication assessment remains a complex process (Vollenweider, 1974; Arhonditsis et al., 2003; Kitsiou and Karydis, 2011) often associated with contrasting physicochemical and bio- logical criteria, spatial heterogeneity, seasonal variability, local conditions, and stochastic processes (Spatharis et al., 2007a). In this frame, phytoplankton biomass (in terms of chl a) remains one of the most commonly used proxies (Karydis and Tsirtsis, 1996), being a simple, direct, and reasonable indicator of eutrophication (Vollenweider, 1974). Numerous approaches have been used for modeling chl a (Kitsiou and Karydis, 2011). Two of the most traditional statistical approaches are linear regression models (Onderka, 2007; Cho et al., 2009) and principal component analysis (Camdevyren et al., 2005; Liu et al., 2010; Primpas et al., 2010). Bayesian statistics have also been applied for chl a prediction using a probabilistic, rather than a simple deterministic approach (Borsuk et al., 2004; Freeman et al., 2009; Ramin et al., 2011). More elaborate approaches include coupled models that incorporate both hydrodynamic and ecological processes (Allen et al., 2007; Lewis and Allen, 2009; Wu et al., 2009). Recent studies have shown that computational tools including fuzzy modeling (Marsili-Libelli, 2004; Pereira et al., 2009), genetic algo- rithms (Muttil and Lee, 2005; Sivapragasam et al., 2010), and ML algorithms (Kompare et al., 1994; Jeong et al., 2006, 2008; Palani et al., 2008; Gonzalez Vilas et al., 2011) can be viable alternatives for modeling eutrophication. * Corresponding author. E-mail address: atamvaki@mar.aegean.gr (A. Tamvakis). Contents lists available at SciVerse ScienceDirect Estuarine, Coastal and Shelf Science journal homepage: www.elsevier.com/locate/ecss 0272-7714/$ e see front matter Ó 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ecss.2012.09.003 Estuarine, Coastal and Shelf Science 115 (2012) 210e217