Empirical modelling of benthic species distribution, abundance, and diversity in the Baltic Sea: evaluating the scope for predictive mapping using different modelling approaches

Abstract

Bučas, M., Bergström, U., Downie, A-L., Sundblad, G., Gullström, M., von Numers, M., Šiaulys, A., and Lindegarth, M. 2013. Empirical modelling of benthic species distribution, abundance, and diversity in the Baltic Sea: evaluating the scope for predictive mapping using different modelling approaches. – ICES Journal of Marine Science, 70: 1233–1243. The predictive performance of distribution models of common benthic species in the Baltic Sea was compared using four non-linear methods: generalized additive models (GAMs), multivariate adaptive regression splines, random forest (RF), and maximum entropy modelling (MAXENT). The effects of data traits were also tested. In total, 292 occurrence models and 204 quantitative (abundance and diversity) models were assessed. The main conclusions are that (i) the spatial distribution, abundance, and diversity of benthic species in the Baltic Sea can be successfully predicted using several non-linear predictive modelling techniques; (ii) RF was the most accurate method for both models, closely followed by GAM and MAXENT; (iii) correlation coefficients of predictive performance among the modelling techniques were relatively low, suggesting that the performance of methods is related to specific responses; (iv) the differences in predictive performance among the modelling methods could only partly be explained by data traits; (v) the response prevalence was the most important explanatory variable for predictive accuracy of GAM and MAXENT on occurrence data; (vi) RF on the occurrence data was the only method sensitive to sampling density; (vii) a higher predictive accuracy of abundance models could be achieved by reducing variance in the response data and increasing the sample size.

Spatial prediction of demersal fish diversity in the Baltic Sea: comparison of machine learning and regression-based techniques

“…However, in our case study, ML methods, especially tree-based ensemble models, surpassed regression-based techniques in terms of prediction accuracy. Similarly, results of studies by Bucas et al (2013) pointed out RF as the most accurate method, when comparing with GAM and MAXENT. In general, the differences in prediction errors obtained for each model by cross-validation were statistically significant.…”

“…Joy and Death, 2004;Bucas et al, 2013;Lopatin et al, 2016). The statistical approaches used for the purpose include more traditional regression-based techniques, such generalized linear models, generalized additive models, and multivariate adaptive regression splines, as well as novel machine learning (ML) algorithms, such as support vector machines, boosted regression trees and random forests.…”

“…Most available predictive applications of ML in marine fish ecology refer mainly to coral reef-associated assemblages in shallow waters (Pittman et al, 2007;Moore et al, 2009;Knudby et al, 2010a;Pittman and Brown, 2011). Such works on demersal fish in other ecosystem types are very scarce (Leathwick et al, 2006a;Monk et al, 2010;Compton et al, 2012;Bergström et al, 2013;Bucas et al, 2013).…”

The potential habitats of two submerged macrophytes,Myriophyllum spicatumandHydrilla verticillatain the river ecosystems, South Korea

“…In addition, abiotic conditions (environmental constraints) and biotic interactions (e.g., competition and herbivory) influence species distributions at a local scale (Austin, 2002;Bučas et al, 2013;Chappuis et al, 2014). The occurrence and abundance of submerged macrophytes are influenced by chemical and physical factors, such as water quality, light availability (Dennison et al, 1993), water transparency, water depth (Canfield et al, 1985), channel slope, channel dimensions (O'Hare et al, 2011), and hydrological regime (Franklin et al, 2008).…”

Shoreline development and degradation of coastal fish reproduction habitats

“…Recent decades have, however, seen a rapid development of species distribution modeling methods, allowing researchers and managers to produce predictive maps of the underwater environment and its associated biota (Elith and Leathwick 2009). In the Baltic Sea, several recent research programs have significantly benefited our understanding and knowledge of habitat distributions in general (Al-Hamdani and Reker 2007;Bučas et al 2013;Lindegarth et al 2014) and coastal fish habitats in particular (Härmä et al 2008;Kallasvuo et al 2009;Sundblad et al 2009Sundblad et al , 2011Sundblad et al , 2013Snickars et al 2010;Bergström et al 2013). For instance, using statistical non-linear relationships between life-stage specific occurrence and environmental descriptors, Sundblad et al (2011) used predictive distribution models to map key reproduction habitats of three of the most common species in the Baltic Sea coastal fish community, northern pike (Esox lucius), Eurasian perch (Perca fluviatilis) and roach (Rutilus rutilus).…”