Excessive spatial resolution decreases performance of quantitative models, contrary to expectations from error analyses

Marine Ecology Progress Series

Mats Lindegarth¹  J. Robin Svensson¹  Lisbeth Jonsson¹ 

关键词: Distribution;    Sessile;    Marine;    Model;    Quantitative;    Spatial;    Scale;   

DOI  :  10.3354/meps10307

学科分类：海洋学与技术

来源: Inter-Research

【 摘 要 】

ABSTRACT: Increased focus on predictive aspects of ecology has recently been urged by scientists and policy makers to provide solutions to pressing societal needs. Current challenges include the large knowledge gap on the spatial distribution of marine biodiversity, and its associated goods and services, and the dependence of model performance on spatial resolution. We evaluated the importance of resolution on the predictive power and precision of empirical models of distributions of marine sessile invertebrates and macroalgae along the Swedish west coast. This was done by simulating the limits to prediction, based on 2 independent simulated proportions of biological variables, and comparing these limits to observed models at different resolutions. Simulations showed the highest achievable predictive power (r²) and precision (RMSE) of models at fine resolutions (~1 m). In contrast to the simulations, the performance of quantitative models was better at relatively coarse resolutions (~100 m). Increased model performance at coarse resolutions could not be explained by differences in sampling or spatial variability. Instead, the improvement is likely caused by the mechanistic coupling (direct or indirect) between predictor variables, depth and hard substratum cover and patterns at coarser scales, whereas complex processes, e.g. biological interactions, shape patterns at finer scales. This match between resolution and the scale at which environmental variables operate may differ among systems, which could explain the discrepancy in outcomes between our study and previous studies. Furthermore, we provide an approach for error analysis that identifies contributions of different model components to the total uncertainty, thus facilitating model optimization.

【 授权许可】

Unknown