From In Situ Observations to Satellites: Machine Learning–Based Modelling of Seawater pCO₂ and pH in the Canary Islands

The improvement of remote sensing systems, together with the emergence of new model-fitting algorithms based on machine-learning techniques, has allowed the estimation of the partial pressure of carbon dioxide (pCO_2,sw) and pH (pH_T,sw) in the waters of the Canary Islands (13-19ºW; 27-30ºN). Continuous time series data from moored buoys and Voluntary Observing Ships (VOS) between 2019 and 2024 were used to train and validate the models, providing an observational foundation for the satellite-based estimations. Among all the fitted models, the most powerful one was the bootstrap aggregation (bagging), giving a RMSE of 2.0 µatm (R² > 0.99) for pCO_2,sw and RMSE of 0.002 for pH_T,sw, although the multilinear regression (MLR), neural network (NN) and categorical boosting (catBoost) also have a good predictive performance, with RMSE ranging from 5.4 to 10 µatm for 360 < pCO_2,sw < 481 µatm and from 0.004 and 0.008 for 7.97 < pH_T,sw< 8.07. Using the most reliable model, it was determined that there is an interannual trend of 3.51 ± 0.31 µatm yr^-1 for pCO_2,sw (which surpasses the rate of increase for atmospheric CO₂ of 2.3 µatm yr^-1) and an increase in acidity of -0.003 ± 0.001 pH units yr^-1. Over the 6 years (2019-2024), the rise in the atmospheric CO₂ and the increase in sea surface temperature, which reached 0.2 ºC per year under the influence of the unprecedented 2023 marine head wave, contribute to this important rate. Considering the Canary Islands, the region has moved from a slight CO₂ source of 0.90 Tg CO₂ yr^-1 in 2019 to 4.5 Tg CO₂ yr^-1 in 2024. After 2022, eastern locations that acted as an annual sink of CO₂ switched to acting as a source.