A Neural Network-Based Estimate of the Seasonal to Inter-Annual Variability of the Lake Superior Carbon Cycle

Abstract

Lake Superior emits and absorbs CO₂ with significant seasonal and interannual variability, which complicates efforts to constrain its carbon cycle. While it regains atmospheric CO₂ equilibrium on sub-annual scales, resulting in a sustained rise in observed pCO₂ over the last two decades, significant gaps in observation have prevented examination of variability in its carbon cycle on smaller temporal or spatial scales. We developed a reconstruction of daily mean Lake Superior surface water pCO₂ and CO₂ lake-air flux with a spatial resolution of 0.02° $\times$ 0.02° in order to infer trends and drivers of carbon cycling in Lake Superior on seasonal to interannual scales. A feed-forward neural network was trained and tested on underway pCO₂ measurements spanning ice-free seasons of 2019–2023, yielding a spatially-comprehensive product describing inorganic carbon dynamics over a five-year period. Lake Superior alternated between net annual CO₂ influx and efflux, with values ranging from $-0.30\pm 0.06{\text{Tg}C\text{yr}}^{-1}$ (influx) to $0.06\pm 0.06{\text{Tg}C\text{yr}}^{-1}$ (efflux) and a 5 year mean net annual $-0.14\pm 0.06{\text{Tg}C\text{yr}}^{-1}$. This refinement of Lake Superior's carbon budget juxtaposes the lake's large seasonal and interannual variability against a mean net annual balance of carbon sources and sinks, and opens the door to further applications of machine learning reconstruction of lacustrine biogeochemical cycling.