09f625b6-20c0-4951-b810-a1d6aa2da9ad
55b44ab3-0b69-4bc4-8d4d-0ad2e022005c
Fanny Lecuy
Fanny Lecuy
Ifremer
Ifremer
sextant@ifremer.fr
sextant@ifremer.fr
2024-01-24T16:05:34.709Z
ISO 19115:2003/19139 - SEXTANT
1.0
2
279
0.009
167
0.009
false
4326
4326
Species and habitats - Merlangius merlangus (Whiting) - Eggs - Preferential habitat in January modellised by Generalised Linear Modelling and its uncertainty for IBTS surveys
Species and habitats - Merlangius merlangus (Whiting) - Eggs - Preferential habitat in January modellised by Generalised Linear Modelling and its uncertainty for IBTS surveys
IBTS_species_model
IBTS_species_model
2009-12-31T00:00:00
CHARM_MERNMER_EGG_GLM_MOD_ERR_R
Source CHARM Consortium
Source CHARM Consortium
Modelised adundance of several species eggs or prediction uncertainty.
Modelised adundance of several species eggs or prediction uncertainty.
IFREMER
IFREMER
CHARM consortium
CHARM consortium
Yves Verin
Yves Verin
Ifremer
Ifremer
Laboratoire de Ressources Halieutiques, Ifremer, 150 quai Gambetta BP699
Laboratoire de Ressources Halieutiques, Ifremer, 150 quai Gambetta BP699
Boulogne sur mer
62321
yves.verin@ifremer.fr
yves.verin@ifremer.fr
Sandrine Vaz
Sandrine Vaz
Ifremer
Ifremer
andre.carpentier@ifremer.fr
andre.carpentier@ifremer.fr
https://sextant.ifremer.fr/geonetwork/srv/api/records/09f625b6-20c0-4951-b810-a1d6aa2da9ad/attachments/mernmer_ibts_glm.png
/Biological Environment/Species/Fish Species of Commercial Interest
/Biological Environment/Species/Fish Species of Commercial Interest
/Milieu biologique/Espèces/Espèces d'intérêt halieutique
Thèmes Sextant
Thèmes Sextant
2023-03-16
geonetwork.thesaurus.local.theme.sextant-theme
Species data set
Species data set
Species data set
CHARM
CHARM
CHARM
Species distribution
Species distribution
Répartition des espèces
GEMET - INSPIRE themes, version 1.0
GEMET - INSPIRE themes, version 1.0
2018-07-27T14:39:48
geonetwork.thesaurus.external.theme.httpinspireeceuropaeutheme-theme
ressource halieutique
ressource halieutique
ressource halieutique
external.theme.gemet
external.theme.gemet
Marine Strategy Framework Directive (MSFD)
Marine Strategy Framework Directive (MSFD)
Directive Cadre Stratégie pour le Milieu Marin (DCSMM)
Cadre Réglementaire - SIMM
Cadre Réglementaire - SIMM
2023-03-16
geonetwork.thesaurus.local.theme.simm.reglementaire
D1: Fish and Cephalopods
D1: Fish and Cephalopods
D1: Biodiversité - Céphalopodes
DCSMM : Descripteurs
DCSMM : Descripteurs
2023-03-16
geonetwork.thesaurus.local.theme.dcsmm-descripteur
Channel-North Sea
Channel-North Sea
Sous-regions marines
Sous-regions marines
2020-06-04
geonetwork.thesaurus.local.place.dcsmm.area
research-only
research-only
Has to be cited this way in maps : "Source CHARM Consortium"
Has to be cited this way in maps : "Source CHARM Consortium"
Has to be cited this way in bibliography : "Carpentier A, Martin CS, Vaz S (Eds.), 2009. Channel Habitat Atlas for marine Resource Management, final report / Atlas des habitats des ressources marines de la Manche orientale, rapport final (CHARM phase II). INTERREG 3a Programme, IFREMER, Boulogne-sur-mer, France. 626 pp. & CD-rom"
Has to be cited this way in bibliography : "Carpentier A, Martin CS, Vaz S (Eds.), 2009. Channel Habitat Atlas for marine Resource Management, final report / Atlas des habitats des ressources marines de la Manche orientale, rapport final (CHARM phase II). INTERREG 3a Programme, IFREMER, Boulogne-sur-mer, France. 626 pp. & CD-rom"
1000
biota
oceans
environment
Microsoft Windows XP ; ESRI ArcGIS 9.x
Microsoft Windows XP ; ESRI ArcGIS 9.x
Eastern English Channel
0.00
2.50
49.50
52.00
Ifremer - Centre de Brest
Ifremer - Centre de Brest
sextant@ifremer.fr
sextant@ifremer.fr
https://sextant.ifremer.fr/services/wms/wms_charm
OGC:WMS
CHARM_MERNMER_EGG_GLM_R
CHARM_MERNMER_EGG_GLM_R
Preferential habitat
Preferential habitat
https://sextant.ifremer.fr/services/wms/wms_charm
OGC:WMS
CHARM_MERNMER_EGG_GLM_MOD_ERR_R
CHARM_MERNMER_EGG_GLM_MOD_ERR_R
Model error
Model error
http://www.ifremer.fr/charm/
WWW:LINK
CHARM web site
CHARM web site
COPYFILE
Preferential habitat
Preferential habitat
Preferential habitat
Preferential habitat
COPYFILE
Model error
Model error
Model error
Model error
The survey concerns the Dover Strait to the southern half of the North Sea. Its aim is to estimate fish abundance and distribution and to compute recruitment indices (abundance of juveniles) for the fish species exploited in the North Sea. Since 2006, French IBTS used a continuous fish egg pumping device called CUFES allowing the collection of information on the spatial distribution of winter spawning areas and on the spawning habitat of several important species.
The survey concerns the Dover Strait to the southern half of the North Sea. Its aim is to estimate fish abundance and distribution and to compute recruitment indices (abundance of juveniles) for the fish species exploited in the North Sea. Since 2006, French IBTS used a continuous fish egg pumping device called CUFES allowing the collection of information on the spatial distribution of winter spawning areas and on the spawning habitat of several important species.
GLM describes the mean response of a species abundance or presence probability according to environmental conditions (figure 2). In this type of model, a linear prediction is related to the mean of the response variable through a link function (e.g. identity function for a normally distributed variable, or logit function for binary data). Corresponding habitat models required a two step modelling procedure. The presence probability of the considered species as a function of environmental factors is first modelled using presence-absence data, independently from abundance data. Then, the response in terms of abundance is modelled in case of presence only. The species¿ habitat can finally be predicted by combining the presence-absence model with the model of abundance response in case of presence. This procedure allows circumventing the problem of atypical distribution of count data which include numerous observations with value zero, which is common in species abundance data (Stefánsson, 1996; Barry & Welsh, 2002). In this study, model selection was carried out by initially fitting a complete model including all available explanatory variables (continuous parameters were introduced as second order polynomials, nominal variables as factors and all first order interactions between environmental parameters were considered; note that interactions were not tested for the egg developmental stage). The selected environment predictors were: temperature, salinity, bed shear stress, depth, chlorophyll a concentration (only for the egg stage), fluorescence (only for the larval stage) and seabed sediment type. Although the first six factors were regarded as continuous covariables, sediment type (mud, fine sand, coarse sand, gravel and pebble) was introduced in the model as a categorical factor. The GLM model was optimised through backward selection based on Chi-square or F-test significance tests (Venables & Ripley, 2002). This approach was taken rather than Akaike Information Criterion reduction (or AIC; Akaike, 1974) to be coherent with quantile regression selection procedure which is also based on significance tests. For presence-absence data, binomial modelling with logit link function was chosen to obtain a prediction of the probability of presence of the species considered. For non-null abundance data (i.e. removing zero values), the data was log-transformed to achieve normality, and gaussian modelling with identity link function was used to predict positive density on a log scale. The predicted probability of presence was then multiplied with the positive density prediction, to obtain the final predicted value of abundance (Stefánsson, 1996). Quantile Regression (RQ) belongs to the family of regression approaches that also includes simple li-near and multiple regression (Koenker, 2005). In RQ, any part of the data distribution may be modelled rather than the mean (a quantile q describing the value greater or equal to q% of the observed data, or in other words the upper bound of q% of the observed data). The study of the upper-bound of response data (typically abundance between 0.75 and 0.95 quantiles) as a function of environmental factor (figure 3) allows estimating their limiting effects on a species distribution (Cade et al., 1999; Hiddink & Kaiser, 2005). As for GLM modelling (see §4.1), the selected environment predictors were: temperature, salinity, bed shear stress, depth, chlorophyll a concentration (only for the egg stage) and fluorescence (only for the larval stage) as continuous covariables and seabed sediment type as factor. Model selection with RQ is made complicated by the large number of candidate models that can be estimated over a range of different quantiles (i.e. one model per quantile): in other words, the model selection includes both the selection of explanatory variables and that of the quantile at which there are considered. Model selection was carried out by initially fitting a model to all available explanatory variables (continuous parameters were introduced as second order polynomials, nominal variables as factors and all first order interactions between environmental parameters were considered; note that interactions were not tested for the egg stage). The selection procedure used is that proposed by Vaz et al. (2008). RQ models were estimated at five quantile intervals, from the 75th to the 95th. Using a backwards elimination procedure, significance tests of all polynomials and interactions were performed and the variable associated with the largest average p-value across the five quantiles, contingent on being greater than 0.05, was selected to be removed from the model. The reduced model hence obtained was then re-run across all five quantiles, and additional variables were removed according to the same rule. Main effects were tested only when associated interactions and second order polynomials had been eliminated. The process of backward elimination was stopped when all remaining variables were significant (p < 0.05) at least for one quantile (Vaz et al., 2008), this quantile being selected as the representative quantile. In case the resulting model was found to have all variables significant over more than one quantile, the highest of these quantiles was chosen, as the more representative of the limiting effect imposed by the environmental variables over the species abundance. For each species considered, the equation of the final habitat model was used to recode digital maps of the environmental factors with the predicted abundance (or presence probability) of the species, using the Raster Calculator tool, thereby producing a habitat map. Prior to this and for each survey, digital (raster) maps of the environmental parameters had been limited to the ranges of values observed during the surveys, so as to avoid extrapolating outside the model development bounds. The resulting habitat maps were further centred and standardised, so that the resulting maps ranged between 0 and 1, thereby permitting an easier comparison amongst results from different stage, species or season (notable exceptions are the habitat maps based on binary data, and the larval stage habitat maps). The spatial distribution of the model error ratios was mapped for each model, the value of 1 corresponding to the maximum possible prediction error. The model prediction error can thus be interpreted as a percentage of model uncertainty.
GLM describes the mean response of a species abundance or presence probability according to environmental conditions (figure 2). In this type of model, a linear prediction is related to the mean of the response variable through a link function (e.g. identity function for a normally distributed variable, or logit function for binary data). Corresponding habitat models required a two step modelling procedure. The presence probability of the considered species as a function of environmental factors is first modelled using presence-absence data, independently from abundance data. Then, the response in terms of abundance is modelled in case of presence only. The species¿ habitat can finally be predicted by combining the presence-absence model with the model of abundance response in case of presence. This procedure allows circumventing the problem of atypical distribution of count data which include numerous observations with value zero, which is common in species abundance data (Stefánsson, 1996; Barry & Welsh, 2002). In this study, model selection was carried out by initially fitting a complete model including all available explanatory variables (continuous parameters were introduced as second order polynomials, nominal variables as factors and all first order interactions between environmental parameters were considered; note that interactions were not tested for the egg developmental stage). The selected environment predictors were: temperature, salinity, bed shear stress, depth, chlorophyll a concentration (only for the egg stage), fluorescence (only for the larval stage) and seabed sediment type. Although the first six factors were regarded as continuous covariables, sediment type (mud, fine sand, coarse sand, gravel and pebble) was introduced in the model as a categorical factor. The GLM model was optimised through backward selection based on Chi-square or F-test significance tests (Venables & Ripley, 2002). This approach was taken rather than Akaike Information Criterion reduction (or AIC; Akaike, 1974) to be coherent with quantile regression selection procedure which is also based on significance tests. For presence-absence data, binomial modelling with logit link function was chosen to obtain a prediction of the probability of presence of the species considered. For non-null abundance data (i.e. removing zero values), the data was log-transformed to achieve normality, and gaussian modelling with identity link function was used to predict positive density on a log scale. The predicted probability of presence was then multiplied with the positive density prediction, to obtain the final predicted value of abundance (Stefánsson, 1996). Quantile Regression (RQ) belongs to the family of regression approaches that also includes simple li-near and multiple regression (Koenker, 2005). In RQ, any part of the data distribution may be modelled rather than the mean (a quantile q describing the value greater or equal to q% of the observed data, or in other words the upper bound of q% of the observed data). The study of the upper-bound of response data (typically abundance between 0.75 and 0.95 quantiles) as a function of environmental factor (figure 3) allows estimating their limiting effects on a species distribution (Cade et al., 1999; Hiddink & Kaiser, 2005). As for GLM modelling (see §4.1), the selected environment predictors were: temperature, salinity, bed shear stress, depth, chlorophyll a concentration (only for the egg stage) and fluorescence (only for the larval stage) as continuous covariables and seabed sediment type as factor. Model selection with RQ is made complicated by the large number of candidate models that can be estimated over a range of different quantiles (i.e. one model per quantile): in other words, the model selection includes both the selection of explanatory variables and that of the quantile at which there are considered. Model selection was carried out by initially fitting a model to all available explanatory variables (continuous parameters were introduced as second order polynomials, nominal variables as factors and all first order interactions between environmental parameters were considered; note that interactions were not tested for the egg stage). The selection procedure used is that proposed by Vaz et al. (2008). RQ models were estimated at five quantile intervals, from the 75th to the 95th. Using a backwards elimination procedure, significance tests of all polynomials and interactions were performed and the variable associated with the largest average p-value across the five quantiles, contingent on being greater than 0.05, was selected to be removed from the model. The reduced model hence obtained was then re-run across all five quantiles, and additional variables were removed according to the same rule. Main effects were tested only when associated interactions and second order polynomials had been eliminated. The process of backward elimination was stopped when all remaining variables were significant (p < 0.05) at least for one quantile (Vaz et al., 2008), this quantile being selected as the representative quantile. In case the resulting model was found to have all variables significant over more than one quantile, the highest of these quantiles was chosen, as the more representative of the limiting effect imposed by the environmental variables over the species abundance. For each species considered, the equation of the final habitat model was used to recode digital maps of the environmental factors with the predicted abundance (or presence probability) of the species, using the Raster Calculator tool, thereby producing a habitat map. Prior to this and for each survey, digital (raster) maps of the environmental parameters had been limited to the ranges of values observed during the surveys, so as to avoid extrapolating outside the model development bounds. The resulting habitat maps were further centred and standardised, so that the resulting maps ranged between 0 and 1, thereby permitting an easier comparison amongst results from different stage, species or season (notable exceptions are the habitat maps based on binary data, and the larval stage habitat maps). The spatial distribution of the model error ratios was mapped for each model, the value of 1 corresponding to the maximum possible prediction error. The model prediction error can thus be interpreted as a percentage of model uncertainty.
IBTS survey, scientific survey of IFREMER
IBTS survey, scientific survey of IFREMER