Evaluation and comparison of data-driven and knowledge-supported Bayesian Belief Networks to assess the habitat suitability for alien macroinvertebrates
Introduction
During the last decades invasive alien species (IAS) are considered to be one of the major drivers of the observed changes in ecosystem structure and functioning (Crowl et al., 2008). Their impact on ecosystems and on the services they deliver can result in high socio-economic costs (Pejchar and Mooney, 2009). Freshwaters are especially vulnerable to the impacts of IAS since they frequently suffer from multiple anthropogenic stressors such as habitat degradation, increased shipping, etc. (Sala et al., 2000). The invasion process depends on different steps (Kolar and Lodge, 2001) and identifying vulnerable sites prior to invasion is the first key step towards effectively managing the undesirable effects related to IAS (Peterson and Vieglais, 2001). Assessing the risk of invasion by alien species can help in prevention efforts, which are significantly cheaper than control efforts after an invasion has occurred (Leung et al., 2002). In this respect, species distribution modelling is a useful tool to generate spatially explicit information on the occurrence and predicted future distribution of IAS (Guisan and Thuiller, 2005). Habitat suitability models have been successfully used in risk assessment to predict the future distribution of IAS as well as to determine ‘hotspots’ for invasion (Ficetola et al., 2007; Jiménez-Valverde et al., 2011; Gallardo and Aldridge, 2013, Boets et al., 2013). These models allow analysing the relation between the presence or absence of the species and environmental conditions (e.g. habitat type, climate or water quality) in order to deduct information on their (future) distribution. However, the large variability of species-specific traits related to invasive species and the fact that often data is used of the invaded area where the species only recently established and thus did not reach its maximum distribution make the prediction of biological invasions difficult (Sutherst and Bourne, 2009).
Bayesian Belief Networks (BBNs) are powerful tools for modelling complex ecosystems and offer potentials for decision support in ecosystem management (Adriaenssens et al., 2004, Stewart-Koster et al., 2010, Spence and Jordan, 2013, Landuyt et al., 2014). However, BBNs have only sporadically been used to predict the habitat preference of IAS to support their management (e.g. Peterson et al., 2008, Murray et al., 2012, Smith et al., 2012, Vilizzi et al., 2013). Despite their sporadic use in this domain, BBNs have been proven to be successful in predicting habitat suitability of aquatic and terrestrial species (e.g. Adriaenssens et al., 2004, Smith et al., 2007). A major benefit of BBNs is the opportunity to complement empirical data with inputs from other models and expert knowledge, in a transparent manner (Low Choy et al., 2009). This is especially useful when modelling the habitat suitability of IAS as there is often not enough quantitative or high quality data available on environmental preferences of the species. Other benefits of BBNs include explicit treatment of uncertainties and the ability to deal with low quality data with missing values (Aguilera et al., 2011, Landuyt et al., 2013). In addition, BBNs can be used for both diagnostic and causal system analysis. Therefore, BBNs are useful tools to design decision support tools that incorporate knowledge of multiple experts and that are designed to assess risks associated with management outcomes.
In this paper, we focussed on gammarids (Gammaridae, Crustaceae), a macroinvertebrate family, because of their important position in the aquatic food web. Gammarids generally have an important function in detritus shredding and are a primary food source for fish (Truhlar et al., 2014). Gammaridae are represented by several highly invasive amphipods species, such as, the ‘Killer shrimp’ Dikerogammarus villosus and Gammarus tigrinus. These species are characterised by high reproductive capacities, an omnivorous diet and a high tolerance to fluctuations in environmental conditions (e.g. Statzner et al., 2008). Recent observations on the amphipod fauna in western and central Europe indicate that native amphipod species are rapidly replaced by alien ones (Jazdzewski et al., 2004, Boets et al., 2011a). These changes in gammarid community composition could have consequences for overall ecosystem functioning and energy flow (Piscart et al., 2011). Truhlar et al. (2014) investigated the leaf shredder efficiency of native and alien invasive gammarids under different environmental conditions. Their results suggest that invasion by alien gammarids could lead to ecosystem-level changes in leaf processing, which could alter nutrient dynamics and community assemblages within invaded systems. Next to altering nutrient dynamics, invasive gammarid species are often highly predatory species that can alter native species composition and reduce biodiversity (MacNeil et al., 2013). Several of these alien gammarid species occur in Belgium and are spreading rapidly to other regions in Western Europe such as the United Kingdom (Boets et al., 2011b, Gallardo and Aldridge, 2013, Boets et al., 2013).
This paper aims to evaluate the use of BBNs to analyse and predict the occurrence of alien gammarids in lowland rivers and streams. To assess the added value of incorporating expert knowledge, we compared the model performance of a purely data-driven (DD) model with the performance of a model that integrates expert knowledge and field data. In this knowledge-supported (KS) data-driven model, personal expert judgement and literature data were used to construct the causal network and to discretize continuous data. Field data were used to quantify the probabilistic relations between the model's variables. The model fit of both models was compared and the importance of predictor variables was evaluated. Finally, the possibilities of using BBNs as decision support technique to define habitats at risk of invasion were discussed.
Section snippets
Data collection and pre-processing
The data used in this analysis originates from the Flemish Environment Agency (VMM), which has been monitoring the water quality in Flanders since 1989 and collected biological, physical-chemical and hydro-morphological data of over 4600 sampling locations situated in inland waters (fresh and brackish water) in Flanders (Belgium). The sampled sites were classified into six types: large rivers (Rg), small rivers (Rk), large streams (Bg), small streams (Bk), polder watercourses (P) and stagnant
Model development settings
The optimal model development settings, defined via multiple model development simulations based on field data, are listed in Table 1. Different state numbers did not result in significant differences in model performance. To reduce model complexity, discretization in two states was chosen as optimal setting. Also discretization type did not have a significant effect on model performance. Equal width discretization was chosen to increase the ecological relevance of the model. An overview of the
Habitat suitability for alien gammarids
Our analysis confirms that river type is the most important variable determining the occurrence of alien gammarids (Fig. 3). Alien gammarids have been frequently observed in large rivers and watercourses with elevated conductivities in central and Western Europe (Bij de Vaate et al., 2002, Grabowski et al., 2009). Large rivers and canals have been pinpointed as hotspots for alien aquatic species introduction, dispersal and establishment. These systems are often subjected to anthropogenic
Conclusion
Bayesian Belief Networks (BBNs) proved to be a suitable conceptual modelling technique offering advantages and opportunities in habitat suitability modelling of alien invasive species. The best model fit for the data-driven BBN was obtained when using the Naive Bayes classifier in combination with an equal width discretization of continuous variables. The data-driven BBN model combined with expertise knowledge resulted in a model with a lower technical performance, but with a high ecological
Acknowledgements
The authors wish to thank the Flemish Environment for the opportunity to study their samples. We are also grateful to Rose Sablon and Joost Mertens who helped consulting the collections. We would like to thank Koen Lock, Elina Bennetsen, Wim Gabriels, Sacha Gobeyn and Ans Mouton for sharing their knowledge with regard to the development of the structure used in the knowledge-supported model. Dries Landuyt is supported by a scholarship of the Flemish Institute for Technological Research (VITO).
References (58)
- et al.
Bayesian networks in environmental modelling
Environ. Model. Softw.
(2011) - et al.
Characterising performance of environmental models
Environ. Model. Softw.
(2013) - et al.
Fish recruitment prediction, using robust supervised classification methods
Ecol. Model
(2010) - et al.
Multimetric Macroinvertebrate Index Flanders (MMIF) for biological assessment of rivers and lakes in Flanders (Belgium)
Limnologica
(2010) - et al.
Habitat suitability modelling of rare species using Bayesian networks: model evaluation under limited data
Ecol. Model
(2015) - et al.
Progress in invasion biology: predicting invaders
Trends Ecol. Evol.
(2001) - et al.
A review of Bayesian belief networks in ecosystem service modelling
Environ. Model. Softw.
(2013) - et al.
An ecosystem service approach to support integrated pond management: a case study using Bayesian belief networks – Highlighting opportunities and risks
J. Environ. Manage
(2014) - et al.
Knowledge-based versus data-driven fuzzy habitat suitability models for river management
Environ. Model. Softw.
(2009) - et al.
Invasive species, ecosystem services and human well-being
Trends Ecol. Evol.
(2009)
Using a Bayesian belief network to predict suitable habitat of an endangered mammal–The Julia Creek dunnart (Sminthopsis douglasi)
Biol. Conserv.
Effects of nitrogen inputs on freshwater wetland ecosystem services–A Bayesian network analysis
J. Environ. Manage
Knowledge discovery from imbalanced and noisy data
Data Knowl. Eng.
Model development of a Bayesian Belief network for managing inundation events for wetland fish
Environ. Model. Softw.
AGIV, Agentschap voor Geografische Informatie Vlaanderen, 2006
Application of Bayesian Belief networks for the prediction of macroinvertebrate taxa in rivers
Ann. Limnol. – Int. J. Limnol.
Geographical patterns in range extension of Ponto–Caspian macro-invertebrate species in Europe
Can. J. Fish. Aquat. Sci.
Shifts in the gammarid (Amphipoda) fauna of brackish polder waters in Flanders (Belgium)
J. Crustac. Biol.
Using long-term monitoring to investigate the changes in species composition in the harbour of Ghent (Belgium)
Hydrobiologia
Using an integrated modelling approach for risk assessment of the ‘killer shrimp’ Dikerogammarus villosus
River Res. Appl.
The spread of invasive species and infectious disease as drivers of ecosystem change
Front. Ecol. Environ.
A key to the freshwater Amphipoda (Crustacea) of Germany
Lauterbornia
Model-based evaluation of ecological bank design and management in the scope of the European Water Framework Directive
Ecol. Eng.
Prediction and validation of the potential global distribution of a problematic alien invasive species – the American bullfrog
Div. Distrib.
Analysis of macrobenthic communities in Flanders, Belgium, using a stepwise input variable selection procedure with artificial neural networks
Aquatic Ecology
Application of bioclimatic models coupled with network analysis for risk assessment of the killer shrimp, Dikerogammarus villosus, in Great Britain
Biol. Invasions
Priority setting for invasive species management: integrated risk assessment of multiple Ponto Caspian invasive species into Great Britain
Ecol. Appl.
Salinity-related distribution of alien amphipods in rivers provides refugia for native species
Biol. Invasions
Predicting species distribution: offering more than simple habitat models
Ecol. Lett.
Cited by (22)
Quantification of multiscale links of key factors with watershed nitrogen and sediment exports based on a Bayesian modelling approach
2023, Journal of Cleaner ProductionUsing community science data to help identify threatened species occurrences outside of known ranges
2022, Biological ConservationCitation Excerpt :Table. A1 and Fig. A2 in Appendix A). For other variables, we used equal width discretization as in Boets et al. (2015) (Table. A2). We used Netica 6.08 (Norsys Software Corporation, 1998), one of the most commonly used BN software package among environmental studies (Aguilera et al., 2011) to develop and evaluate BNs.
A Bayesian Belief Network learning tool integrates multi-scale effects of riparian buffers on stream invertebrates
2022, Science of the Total EnvironmentRisk-informed knowledge-based design for road infrastructure in an extreme environment
2021, Knowledge-Based SystemsA conceptual model for simulating responses of freshwater macroinvertebrate assemblages to multiple stressors
2020, Ecological IndicatorsCitation Excerpt :These models seem also promising because they can include expert-based and other knowledge, which would otherwise be lost, for defining the model structure and for formulating the probabilistic relationships which define the dependencies within the model (McCann et al., 2006). Such model structures have been previously applied to macroinvertebrates (Adriaenssens et al., 2004; Boets et al., 2015; Forio et al., 2015; McLaughlin and Reckhow, 2017; Skeffington et al., 2014) and several other biotic endpoints representing ecological water quality (Johns et al., 2017; Marcot et al., 2001). Thus, this model design is a useful building block for a conceptual model for simulating macroinvertebrate responses to multiple stressors.