3.3.2. Caveats and Recommendations

This study is limited to the estimation of some elements of the carbonate chemistry for the mid-latitude of the North Atlantic Ocean based on their variability during the late winter, spring, and autumn of 2015 and 2016. Whether this neural network algorithm is applicable to other regions of the global oceans and/or for other time periods needs further investigation. The further development and training of the ANN algorithm is therefore recommended. This can be carried out by incorporating (1) a larger scalar variability of the same environmental drivers that are used at the highest spatiotemporal resolution possible in order to improve the learning of the BPN model, and (2) new environmental drivers, such as daily air–sea surface heat fluxes, 2 m air temperature, and air pressure at the highest spatial resolution possible. These may include freshwater influx through precipitation and conditions of the air–sea interface, such as heat fluxes (latent and sensible) and related physical values (such as the sub-layer depth [46]). The atmospheric conditions at sea level are an important parameter that influence the solubility of CO2 in a unit volume of liquid [86]. Increasing the range of EO-based environmental drivers is now becoming more technically feasible, thanks to cloud servers and computing. Equally important would be the derivation of *p*CO2 as another predictand from our artificial neural network algorithm [87]. Due to the limited time available in obtaining high resolution atmospheric and ocean modeled data, the inclusion of these additional environmental drivers was beyond the scope of the present study. The incorporation of (3) dynamical adjustments made to numerical ocean models [88] on the basis of chosen environmental drivers may further enhance the accuracy of the BPN algorithm. For example, it is necessary to take time-dependent temperature variations into account whenever the wind stress is estimated since it varies by more than a factor of two between 0◦ and 30 ◦C because of its dependence on temperature (the Schmidt number).

It is expected that the demand for high resolution DIC, TA, and pH maps, as estimated by deep learning, will, for many reasons, increase in the future. One important use is their support in the monitoring of proposed Ocean Acidification Refugia (OAR), such as the likes of extensive seagrass meadows and dense algal beds [89,90], and algal boundary layers [91,92], slow-flow habitats [93], deep-sea mounts [94], and areas that are isolated from ocean upwelling [95,96]. These are examples of highly localized areas that can vary dramatically across spatial scales from few millimeters (in the case of algal boundary layers) to hundreds of meters squared (such as in the case of extensive seagrass beds), with no clear criteria as to what makes each area a potential OAR other than the observed transient increases in seawater pH relative to the surrounding waters. Kapsenberg and Cyronak (2019) point out the lack of clear, agreed-upon functional criteria for OAR in the context of climate change, which makes it difficult for managers, legislators, and scientists to assess where to invest management efforts [97]. In this regard, this study becomes promising as a way to provide a means by which the daily determination of carbonate chemistry can be made available across multiple spatial scales down to at least a 4 km2 horizontal resolution. In doing so, new target refugia can be proposed for research and management purposes.

#### **4. Conclusions**

Changes in ocean carbonate chemistry are a large spatiotemporal scale phenomenon that certainly needs to be monitored at the local scale. This study addresses its first research question by showing a way to produce high resolution, accurate, gridded maps of DIC, TA, and pH that are ideally suited for more localized ocean carbonate studies and applications.

Ship-based sampling remains subjected to limited ship time and human resources, costs, and weather conditions that prevent sampling in specific areas or at certain times of the year. Yet, they remain fundamental for numerical model validation and initialization tasks. This study shows a way to generate very-high-resolution gridded maps of ocean surface DIC, TA, and pH using an ANN approach in a robust and efficient way. This was carried out by addressing the second research question of this study. The future availability of more EO products hosted by cloud-serving computing environments and deep learning will soon be a determining factor towards the future automation of the synthesis of similar, highly detailed, daily carbonate chemistry maps for the global oceans. This technology will definitely help various ocean-related communities to better mitigate and adapt to the expected long-term changes. This is why we feel that high resolution EO products, coupled with deep learning, will provide us with an indirect way to monitor the chemical changes in seawater at an unprecedented resolution.

**Author Contributions:** Conceptualization, C.G.; Data curation, C.G.; Formal analysis, C.G.; Investigation, C.G.; Methodology, C.G.; Resources, C.G.; Software, C.G.; Validation, C.G.; Visualization, C.G.; Writing—original draft, C.G.; Writing—review and editing, C.G. and R.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by COST Action CA 15217 Ocean Governance for Sustainability— Challenges, Options, and the Role of Science through short-term scientific missions (STSM) on the 'Modeling Ocean Acidification in the Gulf of Cadiz (MOsAiGC)' [CA15217-STSM-39764], and 'Contemporary Acidification Trends in the coastal NorThEaStern Atlantic (ATlaNTES)' [CA15217-STSM 40669].

**Institutional Review Board Statement:** IES-2022-00027. University of Malta.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** The authors would like to acknowledge Barbero, Leticia, Wanninkhof, Rik, Pierrot, and Denis for the collection of dissolved inorganic carbon, total alkalinity, pH, and other variables collected from surface and discrete observations using a flow-through pump and other instruments from M/V Equinox in the North Atlantic Ocean from 7 March 2015 to 6 November 2016 (NCEI Accession 0154382). The US DOC, NOAA, OAR, and the Atlantic Oceanographic and Meteorological Laboratory are also being acknowledged for hosting and making available this information: https://www.ncei.noaa.gov/access/metadata/landing-page/bin/iso?id=gov.noaa. nodc:0154382 (accessed on 20 February 2023). We would also like to acknowledge COST Action CA 15217 Ocean Governance for Sustainability—Challenges, Options, and the Role of Science—for supporting short-term scientific missions (STSM) on the 'Modeling Ocean Acidification in the Gulf of

Cadiz (MOsAiGC)' DOI: 10.13140/RG.2.2.32657.99681, and 'Contemporary Acidification Trends in the coastal NorThEaStern Atlantic (ATlaNTES)'. COST is supported by the EU Framework Program Horizon 2020.

**Conflicts of Interest:** The authors declare no conflict of interest.
