**1. Introduction**

The global oceans constitute an important component in the global carbon cycle. They are also a major sink of human-induced emissions of CO2. When CO2 dissolves under typical ocean surface conditions, 90% of this CO2 is formed as HCO3 −, 9% as HCO3 <sup>2</sup>−, and only 1% as undissociated CO2 *(aq)* and H2CO3 [1]. The four important parameters that are needed to understand the ocean carbonic acid system include the dissolved inorganic carbon (DIC), the total alkalinity (TA), the pH, and the *p*CO2 in surface water.

In the past decades, most of our understanding of the ocean carbonate system is derived from in situ observations. Now, thanks to global networking programs, observations have increased widely and consistently, due to ship surveys, the ARGOS project, and mooring and autonomous platforms; furthermore, due to the availability of ever more complex biogeochemical models, the understanding of ocean global and regional carbonate system has advanced considerably. These activities provide accurate, long-term time series *f*CO2 datasets, such as those found in the Surface Ocean CO2 Atlas—SOCAT—[2,3] and the Global Ocean Data Analysis Project (GLODAPv2.2022), consisting of data products of biogeochemical data collected through the chemical analysis of water samples, including TA, DIC, and many others [4]. This information now shows that surface ocean waters show around a 26% increase in concentration of hydrogen ions since 1860, which is equivalent

**Citation:** Galdies, C.; Guerra, R. High Resolution Estimation of Ocean Dissolved Inorganic Carbon, Total Alkalinity and pH Based on Deep Learning. *Water* **2023**, *15*, 1454. https://doi.org/10.3390/w15081454

Academic Editors: Luis Garrote and Alban Kuriqi

Received: 25 February 2023 Accepted: 3 April 2023 Published: 7 April 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

to a drop in pH from 8.2 to 8.1 [5]. This change has been mainly attributed to the rising anthropogenic emissions of CO2 [5].

From a measurement point of view, changes in pH occur on a large spatial scale and can be influenced by different environmental parameters, especially at the local scale. Due to their very nature, direct field measurements are inherently limited in spatial (time series, moored stations) and/or temporal resolution (ship surveys). Earth observation (EO), on the other hand, offers an avenue for expanding observations and analyzing the temporal and spatial variability of the global ocean and its properties. While EO has proved to be a difficult tool for the direct monitoring of seawater pH and its impact on marine organisms, satellite remote sensing can indirectly measure this by providing us with a range of related physico-chemical and biological processes occurring at the ocean surface at an unprecedented spatiotemporal scale. In addition, even though in situ surface measurements offer a geographically limited representation of the entire oceanic volume and its contents, remote sensing observations of the global ocean become very important for the study of the carbonate system, due to the fact that the change in ocean chemistry arises first in the ocean surface. Thus, environmental satellites have great potential in this field.

At the local level, coastal communities are most vulnerable to a lowering pH, especially where the ocean chemistry is changing most rapidly due to multiple stressors. These communities have the potential of being the worst hit, both economically and socially, especially those who derive benefits from calcifying organisms and other vulnerable species [6]. This explains the need for the rapid monitoring of such coastal waters.

This study asks the following research questions: (1) how can we provide information on the state of ocean carbonate information (such as pH and other important carbonate chemistry parameters) at suitable geographical scales that are useful for the management of marine resources? and (2) how can a more robust monitoring of the ocean carbonate system be made available; one that is chemically, biologically, and physically linked to a good number of environmental drivers instead of a much smaller number of parameters, such as salinity, temperature, and chlorophyll? [7].

In seeking to address these research questions, this study moves away from others that have modeled ocean carbonate parameters at coarse temporal [8] and spatial scales (around 500–1500 km; [9]). Instead, it aims to provide ocean carbonate system parameter information at an unmatched high spatial (4 km) and temporal (such as daily) level via gridded ocean maps, with the opportunity of assimilating this into daily operational monitoring and forward the modeling that is used by a wide variety of ocean end users. This goes perfectly in line with NOAA-SOCAN's top research priorities, i.e., "to monitor key ocean parameters across various spatial and temporal scales that will provide information on mechanistic drivers of acidification and input parameters for predictive model algorithm development" (known as 'priority 1 ) by developing "operational and qualitative models that can transition to end users and adapting existing models to understand acidification" (known as 'priority 3 ) from a "regional perspective as well as in specific systems" [10]. The end-user sectors of this data may range from artisanal and small-scale or semi-industrial fisheries and bivalve aquaculture [11] to coastal managers and policy makers whose actions need to become more adaptive in the short term.

To resolve this challenging aim, this study uses the artificial neural network (ANN) method to fix those specific, inter-related environmental conditions that can lead to particular states of the ocean carbonate system. It does so by following the approach that has been taken by the latest ocean research that uses time-finite, individual-ship-based transect measurements that cross extended oceanic areas such as the North Atlantic Ocean [12], the northwest European shelf seas [13], and the North Pacific Ocean [14], among others.

The calculations that have been carried out in this study were performed at a very high spatiotemporal resolution of a so-far unique list of environmental drivers that, in combination, are able to describe and model the much-needed detailed spatiotemporal variation of surface DIC, TA, and pH. This approach can lead to the prediction of a unique set of high-resolution, daily DIC, TA, and pH regional ocean surface grid maps, with

potential applications in future studies focused on the local dynamics of the carbonate systems in both coastal and oceanic areas.

Now, the vast availability of daily EO data and related ancillary data are ideally suited for the ANN's model-free estimators and for predictive data mining. In this study, the ANN allows the processing of different chemical, biological, and physical ocean values by estimating the most probable field values on the basis of their previous patterns, as observed out in the field. Depending on the algorithmic architecture, the ANN is able to perform its estimations through association, clustering, and prediction of the required output variables. While keeping in mind the practicality and the feasibility of this study, it is very important to create an ANN architecture that is able to learn, and ultimately model, the association between the ocean carbonate parameters and the largest possible number of oceanic physicochemical and biological processes. The potential use of such a tool can be extremely important for the validation of numerical ocean modeling and the prediction of changes in ocean carbonate chemistry.
