A global monthly 3D field of seawater pH over 3 decades: a machine learning approach

Abstract. The continuous uptake of anthropogenic CO2 by the ocean leads to ocean acidification, which is an ongoing threat to marine ecosystem. The ocean acidification rate has been globally documented in the surface ocean, but this information is limited below the surface. Here, we present a monthly 4D 1°×1° gridded product of global seawater pH on the total scale and at in situ temperature (without standardization to 25 °C), derived from a machine learning algorithm trained on pH observations from the Global Ocean Data Analysis Project (GLODAP). The proposed pH product covers the years from 1992 to 2020 and depths from the surface to 2 km on 41 levels. A three-step machine-learning-based algorithm was used to construct the pH product, incorporating region division via a self-organizing map neural network, predictor selection via the stepwise regression algorithm that adds and removes variables from network inputs based on their contribution to reducing reconstruction errors, and nonlinear relationship regression by feedforward neural networks (FFNNs). The performance of the machine learning algorithm was validated using real observations with a cross-validation method, in which four repeating iterations were carried out with each iteration utilizing a different 25 % subset of observations for validation and the complementary 75 % subset for training. The proposed pH product is evaluated using comparisons to time-series observations and the GLODAP pH climatology. The overall root-mean-square error between the FFNN-reconstructed pH and the GLODAP measurements is 0.028, ranging from 0.044 at the surface to 0.013 at 2000 m. The pH product is distributed via the Marine Science Data Center of the Chinese Academy of Sciences: https://doi.org/10.12157/IOCAS.20230720.001 (Zhong et al., 2023).