Which environmental variables best predict distributions of Argentina’s forest birds?

OLAH ASHLEY; RADELOFF VOLKER; SILVEIRA, EDUARDA M.O.; MARTINUZZI SEBASTIÁN; MARTINEZ PASTUR GUILLERMO; POLITI NATALIA; RIVERA LUIS; PIDGEON ANNA

College Park

Workshop; NASA Biodiversity and Ecological Forecasting Team Meeting; 2022

NASA

Halting biodiversity declines is a major conservationgoal. This requires understanding environmental correlates of biodiversity patterns.Traditional measures, like land cover classes or climate, may fall short incharacterizing factors that influence species distributions. Recently developedmeasures that characterize phenology and heterogeneity have potential toincrease predictive power of models. Tocompare the power of novel remotely-sensed measures with established traditionally used environmental variables to predict forest bird speciesdistributions in Argentina. Our studyarea are the forested regions of Argentina. We collected data from eBird with data for 25 forest affiliatedbird species. We used novel remotely-sensed environmental variables: modeledforest structure, phenoclusters1, spatio-temporal variability in greenness  and land surface temperature (LST)2. We usedalso ‘Traditional’ environmental variables: elevation (srtm), precipitation(BIOClim), land cover, soil type, ecoregion. Resolution of allenvironmental variables was 1km. We modelled species distributions with 3algorithms: GLM, Maxent, Random Forest and produced 3 models: remotely-sensedvariables only, ‘traditional’variables only; all variables. We used 4-foldcross validation – 70% training, 30% testing, ensemble models with AUC ≥0.8,TSS ≥0.5, compared ensemble models for each species with AUC & TSS, calculatedvariable importance and response curves for each species from the ensemblemodel with greatest AUC & TSS. Ensemble models of species distributionsbased on novel variables, ‘traditional’ variables, or all variables werecompared based on AUCROC and TSS scores. For some species, like Saltatorsimilis, all variables sets resulted in models with similar predictivepower Phenoclusters and spatialheterogeneity in winter land surface temperature were important for mostspecies in models based on novel environmental variables only. Ecoregion,followed by precipitation seasonality and precipitation of the driest quarterwere the three variables of greatest importance for most species in models basedonly on ‘traditional’ variables as well as models based on all variables. For asubset of forest affiliated bird species in Argentina, ecoregion andprecipitation seasonality were major predictors of their geographicaldistributions. Novel remotely-sensed measures, particularly phenoclusters, spatialheterogeneity of winter LST and mean forest height, were important predictorsfor some species.