A New Bottom Water Climatology Using a Stacked Random Forest and Objective Mapping Approach

Paige D Lavin, Cooperative Institute for Satellite Earth System Studies - University of Maryland, Earth System Science Interdisciplinary Center, College Park, United States and Gregory C Johnson, Pacific Marine Environmental Laboratory, Seattle, United States
Monitoring and understanding changes in deep ocean temperature, salinity, and other water properties is critical for quantifying the global energy imbalance, predicting future sea level rise, and improving climate models. While analyzing such changes in bottom waters helps improve our understanding of the ocean’s role in mediating climate change, it is also important to have a clear picture of the mean state of these abyssal waters. The most recent global synthesis of mean bottom water properties of which we are aware relied on hand-contouring to create maps from sparse shipboard measurements [Mantyla and Reid, 1983]. Since then high-quality sampling of the deep ocean has increased substantially and, concurrently, many advances have been made in the field of machine learning. Hence we have created a new bottom water climatology by using random forests followed by objective mapping to map the sparsely sampled quantities of temperature, salinity, oxygen, silicate, nitrate, and phosphate to a 0.5° smoothed bathymetric grid based on ETOPO2. A random forest is composed of decision trees (each a slightly different set of relationships between the input variables and the output variable) built on many bootstrapped subsamples of the training dataset. For a given set of input variables, the random forest algorithm averages the distinct estimates from all of the decision trees to predict a single output value for those inputs. Our random forest is limited to using latitude, longitude, and depth to predict temperature. However, we do this mapping in an iterative manner, using the new maps of the better sampled fields (e.g., temperature) to help improve the prediction of the more sparsely sampled fields (e.g., silicate). By comparing our new climatology to those produced using other methods more commonly employed by the oceanographic community, we demonstrate the skill of this approach for spatial interpolation of other sparsely sampled oceanographic quantities.