Muriel Gevrey

Review and comparison of methods to study the contribution of variables in artificial neural network models

Abstract Convinced by the predictive quality of artificial neural network (ANN) models in ecology, we have turned our interests to their explanatory capacities. Seven methods which can give the relative contribution and/or the contribution profile of the input factors were compared: (i) the ‘PaD’ (for Partial Derivatives) method consists in a calculation of the partial derivatives of the output according to the input variables; (ii) the ‘Weights’ method is a computation using the connection weights; (iii) the ‘Perturb’ method corresponds to a perturbation of the input variables; (iv) the ‘Profile’ method is a successive variation of one input variable while the others are kept constant at a fixed value; (v) the ‘classical stepwise’ method is an observation of the change in the error value when an adding (forward) or an elimination (backward) step of the input variables is operated; (vi) ‘Improved stepwise a’ uses the same principle as the classical stepwise, but the elimination of the input occurs when the network is trained, the connection weights corresponding to the input variable studied is also eliminated; (vii) ‘Improved stepwise b’ involves the network being trained and fixed step by step, one input variable at its mean value to note the consequences on the error. The data tested in this study concerns the prediction of the density of brown trout spawning redds using habitat characteristics. The PaD method was found to be the most useful as it gave the most complete results, followed by the Profile method that gave the contribution profile of the input variables. The Perturb method allowed a good classification of the input parameters as well as the Weights method that has been simplified but these two methods lack stability. Next came the two improved stepwise methods (a and b) that both gave exactly the same result but the contributions were not sufficiently expressed. Finally, the classical stepwise methods gave the poorest results.

Modelling the factors that influence fish guilds composition using a back-propagation network: Assessment of metrics for indices of biotic integrity

Abstract Fish assemblages are reckoned as indicators of aquatic ecosystem health, which has become a key feature in water quality management. Under this context, guilds of fish are useful for both understanding aquatic community ecology and for giving sound advice to decision makers by means of metrics for indices of biotic integrity. Artificial neural networks have proved useful in modelling fish in rivers and lakes. Hence, this paper presents a back-propagation network (BPN) for modelling fish guilds composition, and to examine the contribution of five environmental descriptors in explaining this composition in the Garonne basin, south west France. We employed presence–absence data and five variables: altitude, distance from the river source, surface of catchment area, annual mean water temperature, and annual mean water flow. We found that BPN performed better for predicting species richness of guilds than multiple regression models. The standardised determination coefficient of observed values against estimated values was used to characterise model performance; it varied between 0.55 and 0.82. Some models showed high variability which was presumably due to spatial heterogeneity, temporal variability or sampling uncertainty. Surface of catchment area and annual mean water flow were the most important environmental descriptors of guilds composition. Both variables imply human influence (i.e. land-use and flow regulation) on certain species which are of interest to environmental managers. Thus, predicting guilds composition with a BPN from landscape variables may be a first step to assess metrics for water quality indices in the Garonne basin.

Small-scale gold mining erodes fish assemblage structure in small neotropical streams

The current gold rush experienced by the Guiana shield is profoundly disturbing freshwater ecosystems. Indeed, streams act as receptors for the water that drains gold mining sites and that contain a high load of sediment and toxicants. We here investigated how gold mining activities affect the structure of fish assemblages in small forest streams in French Guiana. We selected six streams subjected to different types of gold mining impact (reference, former gold mining and currently exploited sites) to measure the impact of gold mining on downstream fish assemblages, but also to determine the resilience of fish assemblages after stopping mining activities. Although overall descriptors of the assemblage such as species richness and fish biomass were not sensitive to gold mining, the fish taxonomic composition was strongly influenced. Furthermore, we showed that the functional structure of fish assemblages was significantly affected by the mining activities favouring smaller and ubiquitous fish at the expense of bigger and habitat specialist species. Even in areas where mining activities had stopped for some time, site resilience was incomplete.

Assessment of stream biological responses under multiple-stress conditions.

Background, aim and scopeDue to the numerous anthropogenic stress factors that affect aquatic ecosystems, a better understanding of the adverse consequences on the biological community of combined pressures is needed to attain the objectives of the European Water Framework Directive. In this study we propose an innovative approach to assess the biological impact of toxicants under field conditions on a large spatial scale.Materials and methodsArtificial Neural Network (ANN) analyses, focusing on impacts at the community level, were carried out to identify the relative importance of environmental and toxic stress factors on the patterns observed in the aquatic invertebrate fauna from the Scheldt basin (Belgium).Results and discussionOur results show that the use of the backpropagation algorithm of the ANN is a promising method to highlight the relationship between environmental pollution and biological responses. This method allows the effects of chemical exposure to be distinguished from the effects caused by other stressors in running waters. Moreover, the use of an overall estimate for toxic pressure in predictive models enables the links between toxicants and community alterations in the field to be clarified. The ANN correctly predicts 74% of samples with an area under the curve of 0.89 and a Cohen’s κ coefficient of 0.64. Organic load, oxygen availability, water temperature and the nitrate concentration appeared important factors in predicting aquatic invertebrate assemblages. On the other hand, toxic pressure did not seem relevant for these assemblages, suggesting that the water quality characteristics were therefore more important than exposure to toxicants in the water phase for the aquatic invertebrate communities in the study area. However, we suggest that the high organic load encountered in the Scheldt basin may lead to an underestimation of the impact of toxicity.

Diatom typology of low-impacted conditions at a multi-regional scale: combined results of multivariate analyses and SOM

Benthic diatoms have been used for decades as indicators of stream water quality and environmental stress. While classification systems and monitoring methods have been developed mostly in Europe, the search for main factors determining assemblages at various scales has been mainly conducted on the American continent. We analysed a selection of 467 diatom records from stream with minimal human impact, from several countries and regions of Western Europe, using different multivariate techniques and artificial neural networks (ANN). The data matrix contained 123 diatom taxa X 23 environmental variables, and covered 35 major catchments. Data processing involved the use of PCA (Principal Component Analysis), DCA (Detrended Correspondence Analysis), CCA (Canonical Correspondence Analysis) and SOM (Self Organizing Maps). Multivariate analyses were useful for identifying the main environmental gradients, and combination of these analyses and SOM enabled to define 10 ecological groups, composed of key indicator taxa. Some of these groups could be identified as corresponding to near-natural conditions, allowing the definition of a biotypology of benthic diatom along a gradient of alkalinity, conductivity, pH – mainly determined by geological features – and a temperature/elevation gradient. Sensitivity analysis and box-plots of environmental variables helped identify the main factors determining stream conditions for these assemblages, and slightly altered conditions or particular situations were easily detected. Several possible bias were identified, either from imbalance among river types in the database, or from taxonomic and identification problems, or from collection of records from various sources. Taxa distribution maps, obtained from the SOM, have been used as a useful mean for representing auto-ecological properties of benthic diatoms and for identifying dual distributions resulting either from errors, from incorrect taxonomic status or from actual ecological differences within a same taxon. On the basis of available information, factors determining diatom assemblages are similar in different regions and even continents, which raises the question of the relevance of the eco-regional approach for this stream community.

Modeling the chemical and toxic water status of the Scheldt basin (Belgium), using aquatic invertebrate assemblages and an advanced modeling method

Self-Organizing Maps have been used on monitoring sites in several Scheldt sub-basins to identify the main aquatic invertebrate assemblages and relate them to the physico-chemical and toxic water status. 12 physico-chemical variables and 2 estimates of toxic risk were available for a dataset made up of a total of 489 records. Two of the five defining clusters reflecting a relatively clean environment were composed by very well diversified functional feeding groups and sensitive taxa. The cleanest assemblage was mainly linked to the sites from the Nete sub-basin. The three other clusters were inversely described with a dominance of oligochaetes and deposit feeders as well as a bad water quality. Such an analysis can be used to support ecological status assessment of rivers and thus might be useful for decision-makers in the evaluation of chemical and toxic water status, as required by the EU Water Framework Directive.

Modeling the impact of landscape types on the distribution of stream fish species

Modifications of the landscape adjoining streams perturb their local habitat and their biological diversity, but little quantitative information is available on land cover classes that influence the fish species individually. Data collected from 191 sites in the Adour–Garonne Basin (France) were analyzed to assess the effects of land cover on the distribution of fish species. A multimodel approach was carried out to predict fish species using land cover classes and to define the most important classes applying a hierarchical filtering based on artificial neural network method and sensitivity analysis. Firstly, using three single-class models, a selection of the land cover subclasses contributing the most was carried out for each fish species and each class. Secondly, multiclass models were built with all the previously selected subclasses to predict each species (n-selected subclass model). Finally, the percentages of contribution for artificial, agricultural, and forest areas obtained for the different m...

Utility of Sensitivity Analysis by Artificial Neural Network Models to Study Patterns of Endemic Fish Species

Artificial Neural Networks (ANNs) have become more and more frequently used in ecology in the last decade, essentially to resolve forecasting problems (Cornuet et al. 1996; Recknagel et al. 1997; Guegan et al. 1998; Clair and Ehrman 1998; Ozesmi and Ozesmi 1998; Maier and Dandy 1999; Laberge et al. 2000).

Prediction of macroinvertebrate diversity of freshwater bodies by adaptive learning algorithms

The natural distribution of organisms is determined primarily by their environmental requirements (Huntley 1999). Thus, understanding community patterns is important to manage target ecosystems. Especially in aquatic ecosystems, communities of benthic macroinvertebrates are important to monitor changes of the target system. Benthic macroinvertebrates constitute a heterogeneous assemblage of animal phyla and consequently it is probable that some members will respond to stresses placed upon them (Hynes 1960, Hawkes 1979). Many are sedentary, which assists in detecting the precise location of pollutant sources, and some have relatively long life histories. They provide both a facility for examining temporal changes and integrating the effects of prolonged exposure to intermittent discharges or variable concentrations of pollutants (Hellawell 1986). Therefore, it is promising to characterize the changes occurring in communities to assess target ecosystems exposed to environmental disturbances. Species richness is an integrative descriptor of the community (Lenat 1988), as it is influenced by a large number of natural environmental factors as well as anthropogenic disturbances (Cummins 1979, Rosenberg and Resh 1993). The disturbances of environmental factors may lead to spatial discontinuities of predictable gradients and losses of taxa (Ward and Stanford 1979). Species richness is known to be sensitive to environment changes in stream ecosystems (Resh and Jackson 1993), and is used as a biological indicator of disturbance. As with species richness, diversity indices decrease under increasing disturbance and stress on the ecosystem. The Shannon-Weaver diversity index (Shannon and Weaver 1949) is commonly used to describe the diversity of a particular community. The index is a function of both the number of species in a sample and the distribution of individuals among those species (Klemm et al. 1990). The diversity index is often used as an ecological indicator for the assessments of ecosystems (Bahls et al. 1992). Development of methods for patterning spatial and/or temporal changes in communities has currently become an important issue in ecosystem management. The River Invertebrate Prediction And Classification System (RIVPACS) was developed to assess water quality. The RIVPACS and its derivates belong to the first integrated ecological assessment analysis techniques (Wright et al. 1993b, Norris 1995). The models are based on a stepwise progression of multivariate and univariate analyses (Barbour et al. 1999). With nonlinear and complex ecological data, however, nonlinear analysing methods should be preferred (Blayo and Demartines 1991). An artificial neural network is a versatile tool for dealing with problems to extract information out of complex, nonlinear data, and could be effectively applicable to classification and association (Lek and Guegan 2000, Recknagel 2003).