Seite - 124 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

ENVIREMData ENVIronmental Rasters for Ecologicalmodeling (ENVIREM) is an open database of climatic and topographic variables used in species distributionmod- eling and other applications. The database contains 41 variables including aridity and potentialevapotranspiration[12].Anexampleof rasters fromtheENVIREMdatabase is represented in thefigure2. Figure2. AnnualMeanTemperatureDistribution 2.3. Approach The approachwe used in this work is FIGS,which is amethod based on two distinct pathways.Thefirst pathway is usingfilteringwhennoevaluationdata is available.This approachmimics the adaptation patterns of a trait and applies same selection pressure exerted onplants by evolution to develop abest subset containing accessionswith high probabilityofhavingtheadaptive traits.Thesecondpathwayis themachine learningap- proachusedwhenpartial evaluationof thecollection is available.Themachine learning algorithmsfinda function that links adaptive traits, environments (andassociated selec- tionpressures)withgenebankaccessions. In the modeling, the following machine learning algorithms were used: K-nearest neighbours(KNN)[13], Support Vector Machines(SVM)[14], Random Forest(RF)[15], ArtificialNeuralNetworks(NNET)[16] andBaggedCarts(BCART)[17]. Eachmachine learningmodelwas tuned to select best tuningparameters using a training set and then thebestmodelwasselectedbetweendifferentmachine learningmodelsbasedonseveral metrics includingaccuracy, specificity andKappa.Thesemetricswere computedon the test set. Then each traitwas predicted for non-evaluated ICARDAbarley accessions andweas- Z.Azoughetal. /PredictiveCharacterizationof ICARDAGenebankBarleyAccessions124