Page - 122 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

the total collection.Core collections [1],[2] should bedynamic andneed to be adjusted when additional germplasmand new information become available. The remaining ac- cessions in the collection should however still be conserved as a secondary source of diversity.Concerns about core collections include rendering the reservecollectionmore vulnerable to loss, lackofrepresentationofrare,endemicalleles,andpoorrelation to the specific needs of users. To address the latter concern, specialized core collections have beenestablishedaroundaparticular trait, region,or typeofmaterial. For adaptive traits, core and mini-core collections may not capture the needed diver- sity [3].An alternative to randomselection and core collections, the use of the focused identification of the germplasm strategy (FIGS), which is a trait-based approach, as- sists genebankmanagers identify desired geneticmaterialwith highprobability of hav- ing the sought trait. In the last 10 years, ICARDA in collaboration with Vavilov in- stitute in Russia and GRDC-Australia have invested in the development of FIGS that usesgermplasmcollectionsiteagro-climaticandedaphic informationtopredictadaptive traits.Thepremisebehind this approach is that theenvironmentunderwhichwildmate- rialand landraceswilldrive theevolutionandselectionofadaptive traits thatcouldbeof use toplantbreeders. It seeks todetermineandquantify relationshipsbetweencollection siteagro-climaticconditionsandthepresenceofspecifictraits, suchasdiseaseresistance or heat resistance. FIGShasbeen successfully used to identify sources of resistance for several useful traits for breeding globally such as Sunn pest inwheat in Syria, Russian wheataphid inbreadwheat [4], abioticstresses, suchasdroughtadaptation inVicia faba L. [5], resistance tostemrust inbreadanddurumwheat in [6]and[7], andstemrustand stripe rust in accessions ofwheat landraces in [8] and [9]. FIGS is also as an efficient toolof linkinggenebankaccessions toa trait of interest [10]. In this paper, we present a work of a predictive characterization on for ICARDA barleycollectionbuildingon theFIGSapproachbymeansof: 1. Assessingmachinelearningpredictabilityforbarleycollection’scharacterization traits 2. Using themodeling outcomes tomake a predictive characterization of the en- tireICARDAbarleycollectionbyassigningprobabilities tonon-evaluatedacces- sions. 2. MaterialsandMethods 2.1. DatasetsDescription:AccessionsandTraits ICARDA accessions database contains more than 32000 barley accessions including around2400wild relatives, distributedworldwidebut collectedmainly from theFertile Crescent,NorthAfrica,Ethiopia,EastEuropeandSouthEastAsia(seeFig.1). ICARDA barley collection is ranked the second globally and represents 18% of the barley ac- cessions conserved worldwide. More than 40 traits are used at ICARDA, as part of thegenebankconservationeffort, tocharacterizebarleyaccessions includingphenology, growth habit,morphology, yield components and somediseases. In this study,weused eight characterization traits as presented inTable1.Table1 showedadescriptionof the traits thatwe are using formodeling in this study. The number of accessions evaluated is however greater than the number of geographic sites aswe havemultiple accessions Z.Azoughetal. /PredictiveCharacterizationof ICARDAGenebankBarleyAccessions122