Multivariate Spectral Clustering for the Delineation of Ecoclimatic Regions of Turkey

Although temperature and precipitation regimes (trends) are well-studied for Turkey using global/regional climate model results and station time-series, a study that integrates topographic, and edaphic attributes with climatic attributes derived from RCM outputs and GCM-based projections for the delineation of ecoclimatic regions for Turkey has not been done until now. We present multivariate representation of ecoclimatic variables up to 28 and 50 distinct ecoclimatic regions (at 9 km2 resolution) for Turkey and map their historical locations and their projected locations in the future under A2 emission scenario. We have performed a multivariate geographical clustering analysis involving an eigen gap analysis to estimate optimal cluster sizes, through the use of a spectral clustering algorithm (Ng et al., 2002; Von Luxburg, 2007) to quantitatively define ecoregions for a past reference period (1961-1990) and for a 21st Century. The clustering algorithm has been implemented in R and ran on a 128-cores SMP hardware of the National High Performance Computing Center of Turkey. Outputs of global climate model ECHAM5 of Max Planck Institute for Meteorology and NCEP/NCAR reanalysis data (NNRP) were dynamically downscaled to a 27 km resolution using a regional climate model, RegCM3, have been used to generate climatic input layers. For the historical period (1961-1990) 30-year daily climatology of the downscaled NNRP and for the 21st mid-century (2041-2070) downscaled general circulation model ECHAM5 under A2 emission scenario projections have been considered. 9 km resolution maps of daily average, maximum, minimum temperatures and solar radiation have been produced using NCAR Command Language's (NCL) bilinear interpolation method considering also the adiabatic lapse rate correction for the temperature values. Precipitation maps have been split into nine grid points and retain their values. In addition to the calculation of four unique climatic variables (degree-day heat sum during growing season, mean precipitation during growing season, mean solar radiation during growing season, degree-day cold sum during non-growing season); 19 bioclimatic variables adopted from WorldClim-Bioclim have also been generated by this study. Topographic and edaphic variables were resampled to 9-km resolution grid maps from HYDRO1k dataset and Harmonized World Soil Database v.1.2 respectively. All the computations were done in R and input layers were prepared using ArcGIS/ArcMap and the data visualizations were done using NCL. We have compared the 20th C and 21st mid-century spatial distributions of ecoclimatic regions. Based on the changes and geographic shifts in their distributions that highlight vulnerable places, long-term ecological observatory sites have been proposed. This study, will in the long run, provide an explicit framework for an 'ecological observatory network' design for Turkey.