The "Ocean Carbon States" Database: a proof-of-concept application of cluster analysis in the ocean carbon cycle

In this paper, we present a database of the basic regimes of carbon cycle in the ocean as obtained using data mining and patter recognition techniques in observational as well as model data. Advanced data mining techniques are becoming widely used in Climate and Earth Sciences with the purpose of extracting new meaningful information from large and complex datasets. Such techniques need to be rigorously tested, however, in simple, well-understood cases to better assess their utility. This is particularly important for studies of the global carbon cycle, where the interaction of physical and biogeochemical drivers confounds our ability to accurately describe, understand, and predict CO2 concentrations and their changes in the major planetary carbon reservoirs. In addition to observational data of the carbon cycle, numerical simulations of the Earth System are becoming increasingly more complex and harder to evaluate. Without reliable numerical models, however, our predictions of future climate change are haphazard. Here we describe the use of a specific data-mining technique, cluster analysis, as a means of identifying and comparing spatial and temporal patterns extracted from observational and model datasets. As the observational data is organized into various regimes, which we will call "ocean carbon states", we gain insight into the physical and/or biogeochemical processes controlling the ocean carbon cycle in nature as well as how well these processes are simulated by the model. Assessment of cluster analysis results demonstrates that this technique effectively produces realistic, dynamic regimes that can be associated with specific processes at different temporal scales for both observations and the model. Furthermore, these regimes can be used to illustrate and characterize the model biases in the model air-sea flux of CO2 which are then attributed to model misrepresentations of salinity, sea surface temperature, wind speed, and nitrate. The goal of this proof-of-concept study is to establish a methodology for implementing and interpreting k-means cluster analysis on observations and model output which will enable us to subsequently apply the analysis to larger, higher frequency datasets of the ocean carbon cycle. To enable further testing and extending the method discussed, all data and analysis scripts are freely available at data.giss.nasa.gov/oceans/carbonstates/ (DOI: https://doi.org/doi:10.5281/zenodo.996891).