Classification of Reynolds phytoplankton functional groups using individual traits and machine learning techniques

KRUK C.; DEVERCELLI M.; HUSZAR V.L.; HERNÁNDEZ E.; BEAMUD G.; DIAZ M.; SILVA L.H.; SEGURA A.

FRESHWATER BIOLOGY (PRINT)

WILEY-BLACKWELL PUBLISHING, INC

Lugar: Londres; Año: 2017

0046-5070

The Reynolds Functional Groups (RFG) classification scheme is an informativeand widely used method in ecological studies of freshwater phytoplankton. Itclusters species with similar traits, as well as common environmental sensitivitiesand tolerances. However, researchers face the difficulty to classify species intoRFG because it relies in expert opinion, taxonomical knowledge and environmentalinformation, which are not always accessible. Thus, a step forward is to buildgeneral statistical models to classify species into RFG.2. Under the hypothesis that an organism?s response to environmentalconditions determines their functional traits, here represented by the RFG, wepredict that morphology and classification into broad taxonomic groupswill explain RFG independently from environmental information and expertknowledge.3. To evaluate the predictive ability of morphological traits (e.g. volume) and taxonomicaffiliation (e.g. chroococcal Cyanobacteria) as discriminant variables ofRFG, we compiled 1,300 species (264 waterbodies) and applied Random Forest(RF) and Classification and Regression Trees (CART). We divided the data to trainthe models and test their performance.4. RF successfully classified species into the 28 RFG (only c. 10% test error) withan average individual RFG success rate of 84.6 (range = 33%?100%). This is arelatively high percentage of success from an ecological point of view. It suggeststhat the selected variables are able to reconstruct the RFG and representwell environmental preferences, without including information about local environmentalconditions as classifiers, information that was available for the classificationby experts.5. Our results reinforce the functional basis of the RFG as defined by Reynolds(Reynolds et al., 2002, J. Plankton Res. 24, 417?428) and support bothmorphological traits and taxonomic classification as good proxies of phytoplanktonresponses to environmental conditions. A dichotomous key based inthe CART was constructed and an R code to classify species into the RFG isfreely available. This work may help users to classify species into the RFG,including those that were not previously listed in Reynolds classificationsystem.