Rediscovering the Species in Community-Wide Predictive Modeling

Abstract

Broadening the scope of conservation efforts to protect entire communities provides several advantages over the current species-specific focus, yet ecologists have been hampered by the fact that predictive modeling of multiple species is not directly amenable to traditional statistical approaches. Perhaps the greatest hurdle in community-wide modeling is that communities are composed of both co-occurring groups of species and species arranged independently along environmental gradients. Therefore, commonly used "short-cut" methods such as the modeling of so-called "assemblage types" are problematic. Our study demonstrates the utility of a multiresponse artificial neural network (MANN) to model entire community membership in an integrative yet species-specific manner. We compare MANN to two traditional approaches used to predict community composition: (1) a species-by-species approach using logistic regression analysis (LOG) and (2) a "classification-then-modeling" approach in which sites are classified into assemblage "types" (here we used two-way indicator species analysis and multiple discriminant analysis [MDA]). For freshwater fish assemblages of the North Island, New Zealand, we found that the MANN outperformed all other methods for predicting community composition based on multiscaled descriptors of the environment. The simple-matching coefficient comparing predicted and actual species composition was, on average, greatest for the MANN (91%), followed by MDA (85%), and LOG (83%). Mean Jaccard's similarity (emphasizing model performance for predicting species' presence) for the MANN (66%) exceeded both LOG (47%) and MDA (46%). The MANN also correctly predicted community composition (i.e., a significant proportion of the species membership based on a randomization procedure) for 82% of the study sites compared to 54% (MDA) and 49% (LOG), resulting in the MANN correctly predicting community composition in a total of 311 sites and an additional 117 sites (n = 379), on average, compared to LOG and MDA. The MANN also provided valuable explanatory power by simultaneously quantifying the nature of the relationships between the environment and both individual species and the entire community (composition and richness), which is not readily available from traditional approaches. We discuss how the MANN approach provides a powerful quantitative tool for conservation planning and highlight its potential for biomonitoring programs that currently depend on modeling discrete assemblage types to assess aquatic ecosystem health.

Machine Learning Methods Without Tears: A Primer for Ecologists

“…Previous applications include the modeling of species distributions (Mastrorillo et al 1997;Ö zesmi and Ö zesmi 1999), species diversity (Guégan et al 1998;Brosse et al 2001;Olden et al 2006b), community composition (Olden et al 2006a), and aquatic primary and secondary production (Scardi and Harding Results from the regression tree for predicting fish species richness as a function of environmental characteristics for 8236 north-temperate lakes in Ontario, Canada. (A) 10-fold cross-validation (solid circles) and resubstitution (empty circles) relative error for the regression tree.…”

“…ANNs also provide a much more flexible way of modeling ecological data. Model complexity can be varied by altering the transfer function or the inner architecture of the network through an increase in the number of hidden neurons or layers to enhance data fitting, or by increasing the number of output neurons to model multiple ecological response variables, such as multiple species (e.g., Ö zesmi and Ö zesmi 1999) or entire communities (e.g., Olden 2003;Olden et al 2006a). It is this flexibility that has likely led to the increased popularity of neural networks in ecology.…”

The Use of Artificial Neural Networks (ANNs) in Aquatic Ecology

“…Joy & Death (2004) predicted the occurrence of 14 species of fish and crustaceans taking into account up to 31 driving variables. Similarly, Olden (2003) predicted the occurrence of 27 fishes considering 9 physical variables, whereas Olden et al (2006) used 24 variables to infer the occurrence of 16 fish species. Other applications in aquatic ecology different from the prediction of species abundances or occurrences are reported in this paragraph.…”

A critique of the use of indicator-species scores for identifying thresholds in species responses

“…We hoped such findings would call attention to the need for species-level analysis during community-level investigations, regardless of whether TITAN or other appropriate analytical techniques are used. The surge of recent papers on species distribution modeling for enhancing bioassessment supports this view (e.g., Olden et al 2006, Elith and Leathwick 2009, Esselman and Allan 2011. If a better approach exists for linking community and species-level analyses, we suspect it does not involve fitting a population of candidate threshold regression models to each species distribution within a sampled assemblage.…”