DIAZ Monica Mabel
congresos y reuniones científicas
Finding the basis of quantitative Reynolds trait-based functional groups with Machine Learning.
KRUK, C.; DEVERCELLI, M.; HUSZAR, V.; HERNANDEZ, E.; BEAMUD, S.G.; DIAZ, M.; SILVA, L.; SEGURA, A.
Workshop; 17th Workshop Of The International Association Of Phytoplankton Taxonomy And Ecology (IAP); 2014
International Association of Phytoplakton
The classification of Reynolds trait-separated functional groups (TSFG) is one of the most useful approaches to study phytoplankton ecology in freshwater ecosystems. TSFG (e.g. X 1 , M) summarizes a rich base of information clustering species according to functional criteria. TSFG describes groups? environmental preferences and tolerances linking community structure with ecological mechanisms. It was originally delineated according to species co-occurrence patterns, requires environmental information and to some degree relies in expert opinion to classify species. However, clustered species have similar morphology and physiology. Our hypothesis is that easy to determine traits, including morphology (e.g. volume, mucilage presence) and phylogenetic affiliation, might explain most of this scheme features. Therefore, our goal is to contribute to TSFG trough a quantification of the way in which the organisms are classified using Machine Learning tools. To do so we constructed a data base of species classified into TSFG, sampled in different environmental conditions from tropical to subpolar climates. A total of 270 ecosystems (lakes, reservoirs and rivers), 1311 species and 3376 organisms were included. We used classification trees (CART) and Random Forest to classify the organisms into TSFG using morphological traits and taxonomic classes as discriminant variables. We trained the model with a subset (2/3) of the data and compared model classification with the classification done by experts in a test set (1/3). It was possible to classify with relatively low error (10 - 20%) the organisms into the Reynolds TSFG using the selected traits. From the 29 originally included TBFG, 13 groups were misclassified for diverse reasons. For example for some TSFG environmental information regarding ecosystem trophic state is required (e.g. A, C). These results support the functional basis of the TSFG. The construction of a key based in this model would aid classifying organisms into TSFG assisting their use by non experts.