Premium
Reliability of N flux rates estimated from 15 N enrichment and dilution experiments in aquatic systems
Author(s) -
Elskens Marc,
Baeyens Willy,
Brion Natacha,
De Galan Sandra,
Goeyens Leo,
de Brauwere Anouk
Publication year - 2005
Publication title -
global biogeochemical cycles
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 2.512
H-Index - 187
eISSN - 1944-9224
pISSN - 0886-6236
DOI - 10.1029/2004gb002332
Subject(s) - dilution , flux (metallurgy) , a priori and a posteriori , statistics , reliability (semiconductor) , set (abstract data type) , phytoplankton , scale (ratio) , function (biology) , mathematics , biological system , environmental science , chemistry , ecology , computer science , thermodynamics , physics , biology , power (physics) , organic chemistry , philosophy , epistemology , quantum mechanics , evolutionary biology , nutrient , programming language
This paper investigates the estimation behavior of six increasingly complex 15 N models, for estimating flux rates between phytoplankton and dissolved N pools in aquatic ecosystems. The development of these models over the last 40 years reflects increasing realism in the pools and fluxes that constitute the N cycle. The purpose of this paper is to assess how the model results are influenced by the underlying assumptions. For example, with respect to uptake of 15 N by phytoplankton, is anything gained by assuming that regenerated N become isotopically enriched after the introduction of the 15 N label, or is it just as accurate to assume that no source other than the initial 15 N label contributes to the enrichment signal in phytoplankton? To conduct an objective assessment of the models, we compared them to (1) a set of reference values generated numerically by a process oriented model, and (2) real experimental data. The results show that for a number of 15 N models, properties such as accuracy and precision cannot both be optimized under the same conditions, and a compromise must be struck. Oversimplified models risk bias when their underlying assumptions are violated, but overly complex models can misinterpret part of the random noise as relevant processes. Therefore none of the 15 N model solutions can a priori be rejected, but each should carefully be assessed with hypothesis testing. A backward regression strategy based on a statistical interpretation of the cost function (sum of the weighted least squares residuals) was used to select optimal solution subsets corresponding to a given data set.