Reply on RC1

The authors developed a new ML approach to reconstruct global surface ocean pCO2 that considers an impact of different predictors in different ocean regions. Based on Self-Organizing Map method authors defined 11 biogeochemical provinces. A stepwise FFNN regression algorithm was applied to each of these provinces to establish a set of predictors that are highly responsible for pCO2 variability in considered province. Based on selected predictors and analysis of FFNN size (number of neurons) a monthly 1°x1° surface ocean pCO2 product from 1992 to 2019 was constructed. The results show a good agreement with validation data and independent observations.

belong to all provinces adjacent to the nearest province border. Samples in these grids were involved in the FFNN training process of multiple provinces, but only counted once in the validation." Please could you clarify what you mean by "only counted once in the validation"? Is only an output from one province used in the validation? If yes, how do you chose a province from which you take an output?
Response: Thank you for pointing out this unclear description. Due to the definition of new boundaries, in each province additional samples were added, which was outside the original boundary, referred as 'boundary sample' here. Now each province contains two types of samples: original samples and boundary samples. The boundary samples were only involved in the training process and were not set as validation samples in the province that it was defines as boundary samples. For one sample near the boundary, it is a 'original sample' in only one province and is a 'boundary sample' in other provinces. Thus, the sample was involved in the validation of only one province, and was involved in the training process in other provinces as 'boundary sample'. The text was modified as "To obtain a smoother distribution, we extended the boundaries of all provinces 5 1°×1° grids outside and divided the samples inside and outside the original boundary of each province into 'original sample' and 'boundary sample'. For one sample near the boundary, it is a 'original sample' in only one province and is a 'boundary sample' in other provinces. Thus, the sample was involved in the validation of only one province, and was involved in the training process in other provinces as 'boundary sample'."  Fig.6b?

More explicit figures' captions. Please provide more explicit figures' captions, period of presented results, or results averaged over xxxx-xxxx, what are horizontal lines in
Response: Thanks for your suggestion. The horizontal line was the average pCO2 growth rate over each decade (1992-2000, 2001-2010 and 2011-2019).
Not correct conclusion. On page 15 lines 375-379 authors concluded that the difference between FNN1 and FFNN3 is relatively small, because predictors used in FFNN1 and FFNN3 were related to main drivers of pCO2, such as CHLa, xCO2 and MLD. However, same drivers are used in FFNN2. Thus, it cannot explain why FFNN2 shows higher differences with observations.
Response: Thank you for pointing out this mistake. After reconsidering this issue, I think the application of latitude and longitude as predicators of pCO2 may be the reason why FFNN2 shows higher MAE and other validation groups shows relatively closer results. For example, in the province P10 that latitude and longitude were considered not good predictors by the stepwise FFNN algorithm, the three validation groups show significant closer results than that in other provinces. While in other provinces, latitude and longitude were used as predictors in the FFNN1 and FFNN3, decreasing the MAE and RMSE. The text was corrected as "The MAE and RMSE difference between FFNN1 and FFNN3 in some provinces were relatively small. The reason for higher MAE and RMSE showed by the FFNN2 may be the application of latitudes and longitudes as predicators in both the FFNN1 and FFNN3 but not in the FFNN2. In the province P10, latitudes and longitudes were considered not good predictors by the stepwise FFNN algorithm and the results of three validation groups were extremely close.".