Seite - 125 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

signed probability of having the trait. The variable importance on the training set for all predictorswere extracted from thebestmodel.The importanceof apredictor froma machine learningalgorithm isgenerally calculatedbasedon the increase in themodel’s predictionerror afterpermuting thepredictor. In this study,R languagewasused,caret librarywasusedformachine learning[18], rworldmapforgeographicplotting[19]. 3. Results All studied traits showed high tomedium predictability with amodeling accuracy be- tween 0.757 for productive tillering capacity and 0.968 for kernel covering. Random forest(RF)modelwas selectedas thebestmodel for6 traitswhileBCARTwas selected as thebestmodel for two traits, seeTable2.Kappawashigh for all traits, ranging from medium(between0.5 and0.6) and substantial (more than0.6) showing that our predic- tions are not happening through randompredictions. Sensitivities and specificitieswere also high for all traits validating the high predictability of the studied traits using envi- ronmental characterizationof the landracecollectionsites. Table2. Summaryofmodeling results Trait Bestmodel Accuracy Kappa Sensitivity Specificity Days toheading BCART 0.789 0.575 0.752 0.822 Days tomaturity RF 0.772 0.543 0.784 0.761 Kernelweight BCART 0.847 0.687 0.868 0.819 Productive tilleringcapacity RF 0.757 0.514 0.667 0.847 Kernel covering RF 0.969 0.759 0.711 0.99 Kernel rownumber RF 0.862 0.634 0.935 0.667 Growthclass RF 0.865 0.533 0.484 0.97 Yellowrust RF 0.815 0.561 0.872 0.686 Using the best model from machine learning modeling, the non-evaluated acces- sions for each trait are predicted using class probabilities at the threshold of 0.5. The predicted probabilitieswere plotted as a histogram (see Fig.3) to assess the patterns of classprobabilitiesandshowthepowerof themodeling indistinguishingbetweenclasses ofa trait.Differentpatternswere foundfordifferent traits in separatingbetweenclasses. Growth class and productive tillering capacity showed a clear separation between pre- dicted classeswhile themodels for the remaining traits hadamediumcapacity in sepa- ratingbetweenclasses.Model for kernelweight andyellowrust hadalmost nocapacity in separationbetween trait’s classes.Weextracted thevariable importance fromthebest model foreach trait.Figures4and5showtwoexamples forkernelweightandcovering. Potential evapotranspiration (PET) of the driest month, Precipitation of Driest Month anddistance to rivers showingalso theavailabilityofwaterwere important factors influ- encingwhether an accession has a highor lowkernelweight.On the other hand,mean temperature of wettest quarter and precipitation of warmest quarter were the climatic variables themost influencingwhetherabarleyaccessionhasacoveredornakedkernel. Figure 6 is showing amap of the trait and the predicted characterization for the entire Z.Azoughetal. /PredictiveCharacterizationof ICARDAGenebankBarleyAccessions 125