Revisiting probabilistic neural networks: a comparative study with support vector machines and the microhabitat suitability for the Eastern Iberian chub (Squalius valentinus)

Abstract

Probabilistic Neural Networks (PNNs) and Support Vector Machines (SVMs) are flexible classification techniques suited to render trustworthy species distribution and habitat suitability models. Although several alternatives to improve PNNs' reliability and performance and/or to reduce computational costs exist, PNNs are currently not well recognised as SVMs because the SVMs were compared with standard PNNs. To rule out this idea, the microhabitat suitability for the Eastern Iberian chub (Squalius valentinus Doadrio & Carmona, 2006) was modelled with SVMs and four types of PNNs (homoscedastic, heteroscedastic, cluster and enhanced PNNs); all of them optimised with differential evolution. The fitness function and several performance criteria (correctly classified instances, true skill statistic, specificity and sensitivity) and partial dependence plots were used to assess respectively the performance and reliability of each habitat suitability model. Heteroscedastic and enhanced PNNs achieved the highest performance in every index but specificity. However, these two PNNs rendered ecologically unreliable partial dependence plots. Conversely, homoscedastic and cluster PNNs rendered ecologically reliable partial dependence plots. Thus, Eastern Iberian chub proved to be a eurytopic species, presenting the highest suitability in microhabitats with cover present, low flow velocity (approx. 0.3 m/s), intermediate depth (approx. 0.6 m) and fine gravel (64-256 mm). PNNs outperformed SVMs; thus, based on the results of the cluster PNN, which also showed high values of the performance criteria, we would advocate a combination of approaches (e.g., cluster & heteroscedastic or cluster & enhanced PNNs) to balance the trade-off between accuracy and reliability of habitat suitability models. 1 Introduction Humans have facilitated species extinctions, invasions, increased soil erosion, altered fire frequency and hydrology, and incited profound changes in primary productivity and other key biogeochemical and ecosystems processes (Ellis et al., 2010). Therefore, in the face of this global change, forecasting future ecosystem states, such as future species geographic distributions or land-use patterns, is currently a central priority in biogeographical and ecological sciences (Eberenz et al., 2016; Evans et al., 2016). As a consequence of this priority, scientists, conservationists and managers are repeatedly compelled to confront new problems requiring data analysis (LaDeau et al., 2016). Data analysis is largely classified into two broad categories: unsupervised and supervised (Olden et al., 2008). The former focus on revealing patterns and structures in data (e.g., finding groups of co-occurring species), such as the renowned Self-Organising Maps (SOM) (Kohonen, 1982) or the laureate t-Distributed Stochastic Neighbour Embedding (t-SNE) (Van Der Maaten and Hinton, 2008). Conversely, supervised approaches, such as decision trees (e.g., CART; Breiman et al., 1984) or the Generalised Additive Models (GAMs) (Hastie and Tibshirani, 1990),...

Predictive models of fish microhabitat selection in multiple sites accounting for abundance overdispersion

“…Indeed, for species presenting univariate selection for both hydraulic variables, model quality metrics increased in median from 0.04 to 0.06 for R 2 MF and from 0.30 to 0.38 for Spearman ρ. These results confirm the need to further develop multivariate approach as described in previous studies on microhabitat and/or mesohabitat selection (e.g., Dixon & Vokoun, 2009;Le Coarer, 2007;Muñoz-Mas et al, 2018). In contrast, for some species as B. barbus, the introduction of additive variable did not increase model performance and confirmed the assumption that in some cases univariate models perform better than multivariate ones (Millidine, Malcolm, & Fryer, 2016).…”

“…Among the abundant literature on habitat selection, many statistical models have been used to study microhabitat selection (Ahmadi-Nedushan et al, 2006;Conallin et al, 2010). This includes the simple comparison of microhabitat densities across habitat categories (e.g., habitat suitability curves of Lamouroux et al, 1999;Mouton et al, 2012), generalized linear models (GLMs, e.g., Labonne et al, 2003;Jowett & Davey, 2007), fuzzy-models (e.g., Muñoz-Mas, Martinez-Capel, Schneider, & Mouton, 2012) that compute a weighted average of different models, and more recent machinelearning techniques such as random forests (e.g., Shiroyama & Yoshimura, 2016;Vezza et al, 2014) or neural networks (e.g., Fukuda, 2011;Muñoz-Mas, Fukuda, Portoles, & Martinez-Capel,-2018) that are complex non-parametric classification methods (Guisan & Zimmermann, 2000).…”

Machine learning of large‐scale spatial distributions of wild turkeys with high‐dimensional environmental data

“…or hardwood forests (Miller & Conner, 2005;Miller, Hurst, & Leopold, 1999), whereas males prefer hardwood and pine forests (Miller et al, 1999 (Acevedo et al, 2017). Although we only presented the partial plots of RFs in this study, similar partial plots of SVMs and response curves or plots of MaxEnt can be used to examine the relationship between environmental variables and habitat suitability (Elith et al, 2011;Muñoz-Mas, Fukuda, Pórtoles, & Martinez-Capel, 2018;Phillips et al, 2006). Machine learning is a promising tool for species distribution modeling due to its nonparametric approaches and sparsity to overcome difficulties arising from high dimensions of environmental data and sparse data on occurrence, particularly in rare, threatened or endangered species.…”

Movement patterns of forest elephants (Loxodonta cyclotis Matschie, 1900) in the Odzala‐Kokoua National Park, Republic of Congo

“…The relationship between the most relevant input variables and the daily linear displacement by forest elephants was graphically characterised with partial dependence plots (Friedman, ) adapting the code implemented in the package randomForests (Liaw & Wiener, ). Partial dependence plots are based on the predictions rendered by the CRFs obtained after substituting, one at a time and sequentially, the inspected variable by the different values of the variable (Muñoz‐Mas, Fukuda, Pórtoles, & Martínez‐Capel, ). Then, the resulting predictions are used to depict the effect of the inspected variable over the response variable (i.e., mean effect and/or other statistics) accounting for the effects of the other variables within the model by averaging their effects.…”