Physical Clustering of the World's Oceans (based on data extracted from World Ocean Atlas 2013 version 2).
Data.
The physical regions are based on the observations of the World Ocean Atlas 2013 version 2 (WOA13v2*; https://www.nodc.noaa.gov/OC5/woa13/).
We extracted the decadal annual means for nine variables. These variables included: Temperature (°C), Salinity (unitless), Density (kg/m3), Dissolved Oxygen (ml/l), Apparent Oxygen Utilization (ml/l), Silicate (µmol/l), Phosphate (µmol/l), Density (kg/m^3) and Nitrate (µmol/l). The datasets for Temperature, Salinity and Dissolved oxygen were provided at 0.25° resolution. We therefore reprojected the remaining WOA13v2 datasets to the same projection by making each 1° cell in these datasets at 0.25° resolution, while assigning the original value to the four finer resolution cells.
For the seafloor physical regions we included two additional dataset derived from GEBCO bathymetry data (https://www.gebco.net/). The first dataset was the bathymetry across the seafloor, this layer was re-projected to 0.25° resolution, were the cell values were based on the mean values of the finer scale GEBCO layer. We then computed the slope of depth based on the bathymetry raster using the ‘terrain’ function in the ‘raster’ package.
Analysis.
We generated physical clusters for the globe at the surface (0m), 200m, 1000m and the seafloor. For the surface, 200m and 1000m regions, we extracted the single depth layers from the WOA13v2 datasets and generated a matrix which represented the sites by the variables.
For the seafloor, we had to generate interpolated layers at the seafloor based on the WOA13v2 data. We did this by looking at the mean depth of the bathymetry data and undertaking a tri-linear (cubic) interpolation of the WOA13v2 data at that seafloor depth. We subsequently ran a tri-linear interpolation of the WOA13v2 for each variable and generated maps of seafloor environmental conditions. One these maps were generated we extracted each variable into a seafloor site by seafloor physical variable matrix.
All four site by physical variables datasets (0, 200, 1000 and seafloor) were then scaled in an attempt to centre and normalise the data.
For each of these four datasets we then fitted a k-means clustering model from 2 to 40 clusters and looked at the resulting model loglikelihood, AIC and BIC. We then selected the number of clusters at the point were the the log-likelihood converged (i.e. the point were additional centroids only gave a marginal increase in log-likelihood).
The resulting cluster identity was then assigned to each site and used to generate maps of the physical clusters for each dataset. These rasters were then converted to shapefiles.
Boyer, T.P., J. I. Antonov, O. K. Baranova, C. Coleman, H. E. Garcia, A. Grodsky, D. R. Johnson, R. A. Locarnini, A. V. Mishonov, T.D. O'Brien, C.R. Paver, J.R. Reagan, D. Seidov, I. V. Smolyar, and M. M. Zweng, 2013: World Ocean Database 2013, NOAA Atlas NESDIS 72, S. Levitus, Ed., A. Mishonov, Technical Ed.; Silver Spring, MD, 209 pp., http://doi.org/10.7289/V5NZ85MT
