Global High-Resolution Random Forest Regression Maps of Ocean Heat Content Anomalies Using in Situ and Satellite Data

Abstract

The ocean, with its low albedo and vast thermal inertia, plays key roles in the climate system, including absorbing massive amounts of heat as atmospheric greenhouse gas concentrations rise. While the Argo array of profiling floats has vastly improved sampling of ocean temperature in the upper half of the global ocean volume since the mid-2000s, they are not sufficient in number to resolve eddy scales in the oceans. However, satellite sea-surface temperature (SST) and sea-surface height (SSH) measurements do resolve these scales. Here we use Random Forest regressions to map ocean heat content anomalies (OHCA) using in situ training data from Argo and other sources on a 7-day × ¼° grid with latitude, longitude, time, SSH, and SST as predictors. The maps display substantial patterns on eddy scales, resolving variations of ocean currents and fronts. During the well sampled Argo period, global integrals of these maps reduce noise relative to estimates based on objective mapping of in situ data alone by roughly a factor of three when compared to time series of CERES (satellite data) top-of-the-atmosphere energy flux measurements and improve correlations of anomalies with CERES on annual time scales. Prior to and early on in the Argo period, when in situ data were sparser, global integrals of these maps retain low variance, and do not relax back to a climatological mean, avoiding potential deficiencies of various methods for infilling data-sparse regions with objective maps by exploiting temporal and spatial patterns of OHCA and its correlations with SST and SSH.