A Spatially Explicit Uncertainty Analysis of the Air‐Sea CO₂ Flux From Observations

Abstract

In order to understand the oceans role as a global carbon sink, we must accurately quantify the amount of carbon exchanged at the air‐sea interface. A widely used machine learning neural network product, the SOM‐FFN, uses observations to reconstruct a monthly, 1° × 1° global CO₂ flux estimate. However, uncertainties in neural network and interpolation techniques can be large, especially in seldom‐sampled regions. Here, we present a three‐dimensional (latitude, longitude, time) gridded product for our SOM‐FFN observational data set consisting of uncertainties (pCO₂ mapping, transfer velocity, wind) and biases (pCO₂ mapping). We find that polar regions are dominated by uncertainty from gas exchange transfer velocity, with an average 48.7% contribution. In contrast, for subtropical regions, wind product choice contributes an average 50.0%. Regions with fewer observations correlate with higher uncertainty and biases, illustrating the importance of maintaining and expanding existing measurements.