| Abstract: |
This dataset contains Mapped Observation-Based Oceanic Dissolved Inorganic Carbon Monthly fields from 2004 through 2019. We increase the temporal resolution of the monthly climatology of MOBO-DIC (Keppler et al., 2020a) to resolve fields of DIC from January 2004 through December 2019. MOBO-DIC2004-2019 consists of time-varying, gap-filled mapped fields of DIC on 28 depth levels in the upper 1500 m on a 1°x1° grid, at monthly resolution. The original method for Keppler et al. (2020a) as well as an analysis of the seasonal dynamics of DIC at a global scale can be found in Keppler et al. (2020b). The MOBO-DIC mapping method is an extension and adaptation of the SOM-FFN approach by Landschützer et al. (2013), where the first step is to cluster the ocean into regions of similar physical and biogeochemical properties using self-organizing maps (SOM). In the second step, we run a feed-forward network (FFN) in each SOM-cluster to approximate and apply the statistical relationship between the target data (here: DIC), and better constrained predictor data that are available as mapped global fields. We adapted the SOM-FFN method in several ways compared to the original method by Landschützer et al. (2013), that mapped oceanic surface pCO2. As we map the DIC in the water column, we extend the mapping grid from three dimensions (latitude, longitude, and time), to four (latitude, longitude, time, and depth). As different predictors are available and/or meaningful when mapping DIC in the water column, we also have a different set of predictor data compared to the approach used by Landschützer et al. (2013). To overcome potential biases in the random selection of training and internal validation data, as well as boundary problems associated with the SOM clustering, we use a bootstrapping approach, running the SOM-FFN method 15 times. We use 3 different set-ups for the SOMs and run 5 slightly different FFNs in each of the SOM clusters. We take the mean across this ensemble as our final DIC fields. Due to data availability of the predictors, and different statistical relationships in the upper and deep ocean, we run the method separately for two depth slabs: from the surface to 500m, and from 500m to 1500 m. Thus, there may be small discontinuities at 500 m due to this boundary problem, but they are well within the uncertainties. We calculate the uncertainty based on three components: the prediction uncertainty (the standard deviation across the ensemble, global mean is approx. 7 μmol/kg), the uncertainty associated with the measurements (2.4 μmol/kg), and the uncertainty associated with the representation (16 μmol/kg). We use standard error propagation of these three components to obtain the overall uncertainty of MOBO-DIC2004-2019 (global mean is approximately 18 μmol kg−1). We want to emphasize that the uncertainties in our mapped estimate of DIC are considerably larger than the general uncertainties in direct observations of DIC. Thus, they must be considered in the interpretation of the data. Due to how the mapping method works, MOBO-DIC is most robust when using averages or integrals over large regions. For the full description of the method and its validation, please refer to both the Main Text and the Supporting Information of Keppler et al. (in review.). |