Ans Mouton

Knowledge-based versus data-driven fuzzy habitat suitability models for river management

Aquatic habitat suitability models have increasingly received attention due to their wide management applications. Ecological expert knowledge has been frequently incorporated in such models to link environmental conditions to the quantitative habitat suitability of aquatic species. Since the formalisation of problem-specific human expert knowledge is often difficult and tedious, data-driven machine learning techniques may be helpful to extract knowledge from ecological datasets. In this paper, both expert knowledge-based and data-driven fuzzy habitat suitability models were developed and the performance of these models was compared. For the data-driven models, a hill-climbing optimisation algorithm was applied to derive ecological knowledge from the available data. Based on the available ecological expert knowledge and on biological samples from the Zwalm river basin (Belgium), habitat suitability models were generated for the mayfly Baetis rhodani (Pictet 1843). Data-driven models appeared to outperform expert knowledge-based models substantially, while a step-forward model selection procedure indicated that physical habitat variables adequately described the mayfly habitat suitability in the studied area. This study has important implications on the application of expert knowledge in ecological studies, especially if this knowledge is extrapolated to other areas. The results suggest that data-driven models can complement expert knowledge-based approaches and hence improve model reliability.

APPLICATION OF CLASSIFICATION TREES AND SUPPORT VECTOR MACHINES TO MODEL THE PRESENCE OF MACROINVERTEBRATES IN RIVERS IN VIETNAM

Abstract In the present study, classification trees (CTs) and support vector machines (SVMs) were used to study habitat suitability for 30 macroinvertebrate taxa in the Du river in Northern Vietnam. The presence/absence of the 30 most common macroinvertebrate taxa was modelled based on 21 physical-chemical and structural variables. The predictive performance of the CT and SVM models was assessed based on the percentage of Correctly Classified Instances (CCI) and Cohens kappa statistics. The results of the present study demonstrated that SVMs performed better than CTs. Attribute weighing in SVMs could replace the application of genetic algorithms for input variable selection. By weighing attributes, SVMs provided quantitative correlations between environmental variables and the occurrence of macroinvertebrates and thus allowed better ecological interpretation. SVMs thus proved to have a high potential when applied for decision-making in the context of river restoration and conservation management.

Assessment of brown trout habitat suitability in the Jucar River Basin (SPAIN): Comparison of data-driven approaches with fuzzy-logic models and univariate suitability curves

The implementation of the Water Framework Directive implies the determination of an environmental flow (E-flow) in each running water body. In Spain, many of the minimum flow assessments were determined with the physical habitat simulation system based on univariate habitat suitability curves. Multivariate habitat suitability models, widely applied in habitat assessment, are potentially more accurate than univariate suitability models. This article analyses the microhabitat selection by medium-sized (10-20 cm) brown trout (Salmo trutta fario) in three streams of the Jucar River Basin District (eastern Iberian Peninsula). The data were collected with an equal effort sampling approach. Univariate habitat suitability curves were built with a data-driven process for depth, mean velocity and substrate classes; three types of data-driven fuzzy models were generated with the FISH software: two models of presence-absence and a model of abundance. FISH applies a hill-climbing algorithm to optimize the fuzzy rules. A hydraulic model was calibrated with the tool River-2D in a segment of the Cabriel River (Jucar River Basin). The fuzzy-logic models and three methods to produce a suitability index from the three univariate curves were applied to evaluate the river habitat in the tool CASiMiR©. The comparison of results was based on the spatial arrangement of habitat suitability and the curves of weighted usable area versus discharge. The differences were relevant in different aspects, e.g. in the estimated minimum environmental flow according to the Spanish legal norm for hydrological planning. This work demonstrates the impact of the models selection on the habitat suitability modelling and the assessment of environmental flows, based on an objective data-driven procedure; the conclusions are important for the water management in the Jucar River Basin and other river systems in Europe, where the environmental flows are a keystone for the achievement of the goals established in the European Water Framework Directive.

Random forests to evaluate biotic interactions in fish distribution models

Previous research indicated that high predictive performance in species distribution modelling can be obtained by combining both biotic and abiotic habitat variables. However, models developed for fish often only address physical habitat characteristics, thus omitting potentially important biotic factors. Therefore, we assessed the impact of biotic variables on fish habitat preferences in four selected stretches of the upper Cabriel River (E Spain). The occurrence of Squalius pyrenaicus and Luciobarbus guiraonis was related to environmental variables describing biotic interactions (inferred by relationships among fish abundances) and channel hydro-morphological characteristics. Random Forests (RF) models were trained and then validated using independent datasets. To build RF models, the conditional variable importance was used together with the model improvement ratio technique. The procedure showed effectiveness in identifying a parsimonious set of not correlated variables, which minimize noise and improve model performance in both training and validation phases. Water depth, channel width, fine substrate and water-surface gradient were selected as most important habitat variables for both fish. Results showed clear habitat overlapping between fish species and suggest that competition is not a strong factor in the study area. We modeled fish distribution at the mesohabitat scale using Random Forests (RF).We evaluated the effect of interspecific interactions on fish habitat use.RF models are validated using an independent dataset and showed high performance.Results showed a clear habitat overlapping between fish species.Fish interspecific competition seems to be a negligible factor for habitat use.

Application of Probabilistic Neural Networks to microhabitat suitability modelling for adult brown trout (Salmo trutta L.) in Iberian rivers

Probabilistic Neural Networks (PNN) have been tested for the first time in microhabitat suitability modelling for adult brown trout (Salmo trutta L.). The impact of data prevalence on PNN was studied. The PNN were evaluated in an independent river and the applicability of PNN to assess the environmental flow was analysed. Prevalence did not affect significantly the results. However PNN presented some limitations regarding the output range. Our results agreed previous studies because trout preferred deep microhabitats with medium-to-coarse substrate whereas velocity showed a wider suitable range. The 0.5 prevalence PNN showed similar classificatory capability than the 0.06 prevalence counterpart and the outputs covered the whole feasible range (from 0 to 1), but the 0.06 prevalence PNN showed higher generalisation because it performed better in the evaluation and it allowed a better modulation of the environmental flow. PNN has demonstrated to be a tool to be into consideration.

Prevalence-adjusted optimisation of fuzzy habitat suitability models for aquatic invertebrate and fish species in New Zealand

Abstract For many years, habitat suitability models for aquatic species have been derived from ecological datasets by model optimisation. Previous research showed that optimisation of the predictive model performance did not necessarily lead to ecologically relevant models due to the impact of the dataset prevalence. Therefore, the adjusted average deviation was presented as a performance criterion that allowed incorporation of ecological relevance in the model optimisation process. This paper aims to analyse the relation between the adjusted average deviation (aAD) and the training set prevalence for three species in different New Zealand river systems: caddis flies Aoteapsyche spp., large brown trout Salmo trutta and rainbow trout Oncorhynchus mykiss . The aAD was implemented in a hill-climbing algorithm to optimise a fuzzy species distribution model for each species. Specifically, the hypotheses were tested that (1) similar relations between the aAD and the training set prevalence would be obtained, (2) training based on the aAD would lead to more accurate model predictions than training based on more frequently applied performance criteria such as CCI, and that (3) the final fuzzy model would produce a realistic model of habitat suitability. The approach in this paper may improve the transparency of the model training process and thus the insight into habitat suitability models. Consequently, this paper could lead to ecologically more relevant models and contribute to the implementation of these models in ecosystem management.

Abundance versus presence/absence data for modelling fish habitat preference with a genetic Takagi-Sugeno fuzzy system.

This study compared the accuracy of fuzzy habitat preference models (FHPMs) and habitat preference curves (HPCs) obtained from the FHPMs in order to assess the effect of two types of data [log-transformed fish population density (LOG) and presence–absence (P/A) data] on the habitat preference evaluation of Japanese medaka (Oryzias latipes). Three independent data sets were prepared for each type of data. The results differed according to the data sets and the types of data used. The HPCs showed a similar trend, whilst the degrees of preference were different. The model accuracy also differed according to the data sets used. Although almost no statistical difference was observed, on average, the P/A-based models showed a better performance according to the threshold-independent performance measures, whilst the LOG-based models showed better performance in predicting absence of the fish. These results can be explained partly from the different shapes of HPCs. This case study of Japanese medaka demonstrated the effect of different types of data on habitat preference evaluation. Further studies should build on the present finding and evaluate the effects of data characteristics such as the size of data sets and the prevalence for better understanding and reliable assessment of the habitat for target species.

Species distribution models for sustainable ecosystem management

Abstract Reactions to ongoing loss of biodiversity include a variety of restoration actions and are characterised by high costs and uncertainty. Related decision-making can be supported by developing species distribution models (SDMs) that link predictors (both abiotic and biotic) with biotic response variables (e.g., abundance, occurrence, etc.). SDMs can fill in the gaps of current ecological knowledge and predict the potential impact of environmental (including climate) change on species distributions. As climate change already resulted in species shifting their range and an increased risk of extinction, invasion, and disease propagation, SDMs can act as a valuable tool to estimate future species distributions and their effects on ecosystem functioning and related services. Among the variety of modelling techniques used to predict future species distributions, five modelling techniques are selected: decision trees, generalised linear models, artificial neural networks, fuzzy logic, and Bayesian belief networks. The unique advantages of each modelling technique allow the modeller to choose the most appropriate technique in each particular situation. In turn, each modelling technique is characterised by specific drawbacks and is restricted by the limited ecological knowledge related to biotic interactions. Gathering additional ecological knowledge provides the possibility to go beyond simple pattern recognition and to establish more ecologically sound models.

The distribution of an invasive fish species is highly affected by the presence of native fish species: evidence based on species distribution modelling

Topmouth gudgeon (Pseudorasbora parva) is one of the most invasive aquatic fish species in Europe and causes adverse effects to ecosystem structure and functioning. Knowledge and understanding of the species’ interactions with the environment and with native fish are important to stop and prevent the further spread of the species. Creating species distribution models is a useful technique to determine which factors influence the occurrence and abundance of a species. We applied three different modelling techniques: general additive models, random forests and fuzzy habitat suitability modelling (FHSM) to assess the habitat suitability of topmouth gudgeon. The former two techniques indicated that the abundance of native fish (i.e. biotic variables) was more important than environmental variables when determining the abundance of topmouth gudgeon in Flanders (Belgium). Bitterling (Rhodeus amarus), stone loach (Barbatula barbatula), three-spined stickleback (Gasterosteus aculeatus) and predator abundance were selected as the most important biotic variables and implemented in the FHSM to investigate species interactions. Depending on the preferred food source and spawning behaviour, either coexistence or interspecific competition can occur with bitterling, stone loach and three-spined stickleback. In contrast, the presence of predators clearly had a top down effect on topmouth gudgeon abundance. These findings could be applied as a biological control measure and implemented in conservation strategies in order to reduce the abundance of earlier established populations of topmouth gudgeon.

Integrating data-driven ecological models in an expert-based decision support system for water management in the Du river basin (Vietnam).

In this study, classification trees were combined with the Water Framework Directive (WFD)-Explorer, a modular toolbox that supports integrated water management in a river basin to evaluate the impact of different restoration measures on river ecology. First, the WFD-Explorer toolbox analysed the effect of different restoration options on the abiotic river characteristics based on the water and substance balance embedded in the simulation environment. Based on these abiotic characteristics, the biological index Biological Monitoring Working Party for Vietnam was then predicted by classification trees that were trained on biological and abiotic data collected in the Du river basin in northern Vietnam. The ecological status of streams in the basin ranged from nearly pristine headwaters to severely impacted river stretches. Elimination of point sources from ore extraction and decentralised domestic wastewater treatment proved to be the most effective measures to improve the ecological condition of the Du river basin. The combination of the WFD-Explorer results with data-driven models enabled model application in a situation where expert knowledge was lacking. Consequently, this approach appeared promising for decision support in the context of river restoration and conservation management.