Parameter Optimization of an Ecological Niche Modeling Workflow

Optimization Research Object Summary and Description

The Ecological Niche Modeling (ENM) workflow has been designed to perform the analysis of species distributions and to predict changes in biodiversity patterns [Kulhanek et al 2011, Wiley et al 2003, Guinan et al 2009]. The idea of niche modeling is based on G.E. Hutchinson's definition of the realized niche, where a set of environmental factors or a multidimensional space of resources (e.g. light and structure), can be used to predict the persistence of a species [Hutchinson 1957]. Thus, potential distribution models can be generated with relatively few variables characterizing the abiotic environment of the species in the form of geo-referenced raster layers.

In order to obtain valid models for species niches, in many cases it is important to specify the appropriate set of parameter values for a given input data set (occurrences, mask, and layers). The best parameter values can vary between different input data sets and hence are difficult to know beforehand.

The workflow uses web services to remotely execute a specific modeling algorithm, provided by the openModeller tool. The result model represents the suitable conditions of abiotic for a given species. After creating the model, a test model operation is called. This operation tests the model, by using the 10% points left out of model creation. The test operation calculates the receiver operating characteristic (ROC) curve, area under the curve (AUC). An abstract representation of the workflow can be seen below:

Fig 1: The abstract ENM workflow uses occurrence and environmental data to model ecological niches based on a variety of algorithms, including Maxent, Support Vector Machines and others.

This page represents a Research Object representing the optimizations made on the AUC output parameter of the ENM workflow using Support Vector Machines (SVM). The optimizations are performed using genetic algorithms and have been represented with the Research Object Optimization Ontology. An RDF representation of the Research Object can be downloaded from this link. An alternative representation in RDF-a can be found in this document.

Optimization Research Object: Aggregated Resources

The "Parameter Optimization of an Ecological Niche Modeling Workflow" Research Object aggregates several resources used to create the optimizations of the workflow. This section introduces each of the resources step by step and explains them in detail.

• The ENM workflow being optimized. It has been encoded and executed in Taverna, and it is currently uploaded in myExperiment. The workflow has several input parameters, and some of them have fixed values in all the optimization execution runs. These parameters are: Nu (value=0.5) , Coef0 (value=0.5) , Degree (value=3) , Kernel_Type (value=2) , SVM_Type (value=0) , ProbabilisticOutput (value=1) and Mask (value=wcs>http://biovel.iais.fraunhofer.de/geoserver/ows?>biovel_temp:SJB_Baltic_NorthEastAtlantic _20130522_112604_531) . The parameters being optimized are numberOfPseudoAbstences, Gamma and Cost.
• The Search Space Configuration, with the subworkflow on top of which the search is being performed: the "Select algorithm" subworkflow. The search space configuration for the parameters being optimized is as follows: Gamma has a minimum value 0, a maximum value 10 and round value true. Cost has a minimum value 0, a maximum value 8 and and exponential function with value (base) 2. Finally, numberOfPseudoAbstences has minimum value 200, and maximum value 600.
• A termination condition with maximum time 1440 minutes.
• The Genetic Algorithm used has two output parameters: Cross Over=0.75 and Mutation Rate=0.2 .
• The Fitness function used (single objective), which is performed on the "AUC" parameter.
• A selection of the optimization runs was performed on the Workflow, from more than 100 optimization runs. These runs are specified on the table below (note that they are not in chronological order of execution):

  Optimization Run Number	Cost Value	Gamma Value	NumberOfPseudoAbstences Value	Fitness Value
  Run 1	8	2.36	363	0.9207
  Run 2	64	5.87	458	0.9014
  Run 3	8	6.84	319	0.8974
  Run 4	32	3.79	486	0.9019
  Run 5	256	8.94	568	0.9032
  Run 6	32	0.06	215	0.9084
  Run 7	8	4.78	491	0.8867
  Run 8	8	1.25	562	0.9102
  Run 9	64	6.09	514	0.8929
  Run 10	4	2.24	453	0.9101

References

• [Guinan et al 2009]: J. Guinan, C. Brown, M. F. Dolan, and A. J. Grehan. Ecological niche modelling of the distribution of cold-water coral habitat using underwater remote sensing data. Ecological Informatics, 4(2):83-92, 2009.
• [Hutchinson 1957]: G. E. Hutchinson. Cold Spring Harbor Symposium on Quantitative Biology. Concluding remarks, 22:415-427, 1957.
• [Kulhanek et al 2011]: S. A. Kulhanek, B. Leung, and A. Ricciardi. Using ecological niche models to predict the abundance and impact of invasive species: application to the common carp. Ecological Applications, 21(1):203-213, 2011.
• [Wiley et al 2003]: E. O. Wiley, K. M. McNyset, A. T. Peterson, C. R. Robins, and A. M. Stewart. Niche modeling and geographic range predictions in the marine environment using a machine-learning algorithm. Oceanography, 16(3):120-127, 2003.

Acknowledgements

The authors would like to give special thanks to Olav Zimmermann, Renato De Giovanni, Matthias Obst and Carole Goble for their help with the Research Object. The authors would also like to thank Alan Williams from the BioVEL project for his support on the DR work ow (BioVEL: EU FP7 283359 BioVel BioDiversity eLaboratory). This work was partly supported by the myGrid Platform Grant (EPSRC EP/G026238/1, "myGrid: A platform for e-Biology Renewal"), by the Wf4Ever European project (FP7-270192), and an FPU grant (Formacion de Profesorado Universitario) from the Spanish Science and Innovation Ministry(MICINN).