A novel method for diagnosing seasonal to inter-annual surface ocean carbon dynamics from bottle data using neural networks

[摘要] The ocean's role in modulating the observed 1–7 Pg C yr⁻¹inter-annual variability in atmospheric CO₂ growth rate is an important,but poorly constrained process due to current spatio-temporal limitations inocean carbon measurements. Here, we investigate and develop a non-linearempirical approach to predict inorganic CO₂ concentrations (total carbondioxide (C_T) and total alkalinity (A_T)) in the globalocean mixed layer from hydrographic properties (temperature, salinity,dissolved oxygen and nutrients). The benefit of this approach is that oncethe empirical relationship is established, it can be applied to hydrographicdatasets that have better spatio-temporal coverage, and therefore provide anadditional constraint to diagnose ocean carbon dynamics globally. Previousempirical approaches have employed multiple linear regressions (MLR) andrelied on ad hoc geographic and temporal partitioning of carbon data toconstrain complex global carbon dynamics in the mixed layer. Synthesizing anew global C_T/A_T carbon bottle dataset consisting of~33 000 measurements in the open ocean mixed layer, we develop aneural network based approach to better constrain the non-linear carbonsystem. The approach classifies features in the global biogeochemical datasetbased on their similarity and homogeneity in a self-organizing map (SOM;Kohonen, 1988). After the initial SOM analysis, which includes geographicconstraints, we apply a local linear optimizer to the neural network, whichconsiderably enhances the predictive skill of the new approach. We call thisnew approach SOMLO, or self-organizing multiple linear output. Usingindependent bottle carbon data, we compare a traditional MLR analysis to ourSOMLO approach to capture the spatial C_T and A_Tdistributions. We find the SOMLO approach improves predictive skill globallyby 19% for C_T, with a global capacity to predictC_T to within 10.9 μmol kg⁻¹(9.2 μmol kg⁻¹ for A_T). The non-linear SOMLOapproach is particularly powerful in complex but important regions like theSouthern Ocean, North Atlantic and equatorial Pacific, where residualstandard errors were reduced between 25 and 40% over traditional linearmethods. We further test the SOMLO technique using the Bermuda Atlantictime series (BATS) and Hawaiian ocean time series (HOT) datasets, wherehydrographic data was capable of explaining 90% of the seasonal cycleand inter-annual variability at those multi-decadal time-series stations.

[学科分类] 地球化学与岩石