Machine learning highlights the importance of primary and secondary production in determining habitat for marine fish and macroinvertebrates

Abstract

Species distribution models for marine organisms are increasingly used for a range of applications, including spatial planning, conservation, and fisheries management. These models have been constructed using a variety of mathematical forms and drawing on both physical and biological independent variables; however, what might be called first‐generation models have mainly followed the form of linear models, or smoothing splines, informed by data collected in the context of fish surveys. The performance of different classes of variables were tested in a series of species occurrence models built with machine learning methods, specifically evaluating the potential contribution of lower trophic level data. Random forest models were fitted based on the classification of the absence/presence for fish and macroinvertebrates surveyed on the US Northeast Continental Shelf. The potential variables included physical, primary production, secondary production, and terrain variables. For accepted model fits, six variable importance measures were computed, which collectively showed that physical and secondary production variables make the greatest contribution across all models. In contrast, terrain variables made the least contribution to these models. Multivariable analyses that account for all performance measures reinforce the role of water depth and temperature in defining species presence and absence; however, chlorophyll concentration and some specific zooplankton taxa, such as Metridia lucens and Paracalanus parvus, also make important contributions with strong seasonal variations. Our results suggest that lower trophic level variables, if available, are valuable in the creation of species distribution models for marine organisms.

A systematic review of spatial habitat associations and modeling of marine fish distribution: A guide to predictors, methods, and knowledge gaps

“…in the mid-Atlantic, USA [76,77], which are benthic species that are most common in areas with topographic relief [78]. For demersal fish, sediment grain size was a common predictor; this variable often has a relatively minor influence on species' distribution [19,21], but such data are often limited and have a coarse resolution.…”

“…Recent advances include the Spatial Ecosystem and Population Dynamics Model (SEAPODYM), which is designed to predict pelagic predators based on the predicted distribution of lower and mid-trophic level prey [84,85]. When tested, copepod abundance has been a useful predictor of mackerel [86] and demersal species [19]. Copepod community shifts due to El Niño events have led to a less lipid-rich community than during other years, and therefore, affects the pelagic food chain [87].…”

“…Season-specific metrics of ocean characteristics, such as temperature and salinity, may also be applicable. Digital elevation models may be used to produce metrics of depth, slope, aspect, rugosity, slope, slope of slope, curvature, bathymetric position index, and seabed forms at multiple spatial scales [18,19]. Predictive modeling studies have ranged from those using no predictor variables in favor of kriging geostatistical techniques [20] to those using > 20 different predictor variables [21,22].…”

“…Larger species are also more sensitive to temperature change because of physiological constraints imposed by oxygen capacity and demand [27]. Regarding predictor variable differences, Friedland et al [19] analyzed bottom trawl surveys and found fish functional groups differed in their strength of association with physical, primary productivity, secondary productivity (i.e., zooplankton), and benthic terrain variables. Rather than a one-size-fits-all approach to marine spatial modeling, this information suggests that appropriate predictor variables are likely to differ by fish guild.…”

Automatic Segregation of Pelagic Habitats

“…While it remains difficult to segregate pelagic habitats, which exhibit no clear boundaries (Hinchey et al, 2008;Pittman et al, 2011;Wedding et al, 2011) and can be quite dynamic on a wide range of scales, benthic habitat maps can give an impression of physically distinct areas that consistently occur together with particular species communities (Harris and Baker, 2012). Some effort has been undertaken to characterize fish habitats (e.g., Bellido et al, 2008;Giannoulaki et al, 2011;Tugores et al, 2011;Laman et al, 2017;Amorim et al, 2018;Friedland et al, 2020;Funk et al, 2020), but fewer studies focused on zooplankton (e.g., Labat et al, 2009;Alvarez-Berastegui et al, 2014;Espinasse et al, 2014). Thus, mechanisms contributing to the enormous diversity of plankton, a fundamental component of pelagic food webs, are still not fully understood (Sano et al, 2013;North et al, 2016).…”