Advancing projections of phytoplankton responses to climate change through ensemble modelling

Highlights

• We examined the performance of multiple lake and reservoir ecosystem models.

• The mean of all models was the best predictor of phytoplankton in a temperate lake relative to any individual model.

• Future climate warming scenarios predict increases in total phytoplankton biomass.

• Climate warming will facilitate higher yields of cyanobacteria relative to nutrient levels.

Abstract

A global trend of increasing health hazards associated with proliferation of toxin-producing cyanobacteria makes the ability to project phytoplankton dynamics of paramount importance. Whilst ensemble (multi-)modelling approaches have been used for a number of years to improve the robustness of weather forecasts this approach has until now never been adopted for ecosystem modelling. We show that the average simulated phytoplankton biomass derived from three different aquatic ecosystem models is generally superior to any of the three individual models in describing observed phytoplankton biomass in a typical temperate lake ecosystem, and we simulate a series of climate change projections. While this is the first multi-model ensemble approach applied for some of the most complex aquatic ecosystem models available, we consider it sets a precedent for what will become commonplace methodology in the future, as it enables increased robustness of model projections, and scenario uncertainty estimation due to differences in model structures.

Introduction

Ecological processes in lakes and reservoirs are highly sensitive to environmental change (Williamson et al., 2008). Phytoplankton, being an integral component of lake food webs, has proven to be particularly responsive to changes in factors influenced by global change such as nutrients (Reynolds, 1984, Lampert and Sommer, 1997) and climate (Huber et al., 2008, Jöhnk et al., 2008). This sensitivity is of particular concern to lake managers around the world and, when coupled with the related health hazard of cyanobacterial blooms (Chorus and Bartram, 1999, Paerl and Huisman, 2008), makes understanding their ecology and projecting their biomass of paramount importance.

Projecting cyanobacterial blooms has proved challenging and has motivated the development of numerous computer models that have attempted to simulate the seasonal development of lake phytoplankton (Mooij et al., 2010, Trolle et al., 2012). Nevertheless, with the additional pressures that freshwater ecosystems are subject to with a changing climate, the need for the predictive ability of such models has never been more important than now (Dale et al., 2006, Oliver et al., 2012). However, most studies that simulate future impacts on lake phytoplankton have utilised only a single mechanistic model (Mooij et al., 2007, Trolle et al., 2011, Elliott, 2012). Whilst such studies have merit, the advantage of applying multiple, independently developed models – i.e., an ensemble modelling approach – to a given lake system is that some of the inherent uncertainties in the individual model projections can be reduced by conveying the mean and range of the projections. Nevertheless, no ensemble modelling studies have yet been carried out for projection of lake phytoplankton dynamics, perhaps due to constraints associated with both the considerable resources and the availability of expertise needed to model these aquatic ecosystems.

In this study we take advantage of the expertise available within an international network of modelling experts and apply three autonomously developed models to the same freshwater lake system. We do not intend to provide a comprehensive review of the conceptual differences between these models (for that, we refer to the bulk of literature already available, e.g., Mooij et al., 2010, Trolle et al., 2012), but rather take advantage of these differences in evaluating the diversity of simulated signals they provide. We further simulate a range of future climate change scenarios, represented by a 1.5, 3 and 5 °C warming scenario and two increased nutrient load scenarios, so that the different model projections and likely impacts of warming can be assessed. We test our hypothesis that, as for weather forecast models, the ensemble model mean (derived as the daily average of output from the three individual models) can provide a better predictive working model compared with any individual model (Gneiting and Raftery, 2005.), whilst also allowing statistical uncertainties to be expressed where otherwise they would not be, i.e., with the more common approach of applying a single mechanistic model. We test this method on Lake Engelsholm (Denmark) which, like many other lowland lakes worldwide, has been undergoing eutrophication as a result of decades of anthropogenic impacts from both point and diffuse sources of nutrient pollution.