Response to referee comment 1

Comment:

The variation of Chl-a concentration in the global ocean surface is influenced by various complex factors, which poses challenges to accurately retrieve Chl-a concentration. In this manuscript, the authors selected Chl-a data products retrieved from satellite data as a reference, supplemented by reanalysis data to provide environmental factor information. By combining the advantages of machine learning in big data analysis and simulation, they ultimately reconstructed a global-scale, long-term time series of Chl-a concentration dataset. The results show that the OCNET model performs very well in reconstructing Chl-a concentrations, accurately capturing the temporal variations of these features. This suggests that the model has strong potential for use in large-scale ocean color data reconstruction, and may even be able to predict Chl-a concentration trends in response to changes in the environment.

The dataset is extremely valuable for a wide range of researchers, policymakers, and managers involved in the monitoring and management of aquatic ecosystems. This study and its results are really novel and impressive to me.

Overall, I find this study and its results to be highly promising and valuable for the scientific community. With the suggested clarifications and improvements, this manuscript has the potential to make a significant contribution to the field of Chl-a concentration retrieval and ocean color data reconstruction.

Response:

We really appreciate these overall comments and recommendation by this reviewer. Our point-by-point responses to the reviewer's comments are given as follows.

Specific Comments:

1) Pg. 3, Lines 65-66: Please provide an explanation for the difference between OCNET and the CNNs used in previous studies (Cao et al., 2020; Jin et al., 2021; Cen et al., 2022; Yussof et al., 2021). Clarifying this distinction would enhance the reader's understanding of the novelty of the OCNET model.

Response:

Thanks for this comment. OCNET, a modified U-Net architecture, deviates from the general Convolutional Neural Network (CNN) in terms of network structure. Most data reconstruction methods based on machine learning, including CNNs and random forest, primarily rely on the spatiotemporal correlations within the data, using valuable spatiotemporal sequences to infer missing regions. However, such methods struggle to achieve satisfactory results when dealing with extensive, irregularly distributed missing data. The OCNET machine learning method proposed in this study draws inspiration from oceanic biogeochemical models and incorporates environmental variables that influence the distribution of chlorophyll-a concentration into data reconstruction. By learning the impact of environmental factors on the variation of chlorophyll-a concentration, it enables long-term and large-scale reconstruction of missing data. We will further elaborate on the distinctive features of OCNET in the revised manuscript.

2) Pg. 6, Lines 125-126: The authors state that "We have selected three environmental variables, i.e., sea surface temperature (SST), salinity (SAL), and photosynthetically active radiation (PAR) as the input data for the OCNET model." However, it appears that the input data also contain SSP. Please address this discrepancy and provide clarity on whether SSP is included in the input data or not.

Response:

Thanks for this comment. In fact, we did select four environmental variables, including SSP, as inputs. The three variables mentioned here (SST, SAL, and PAR) are directly related to phytoplankton growth, as further explained in the manuscript. SSP primarily reflects the dynamic environmental factors at the ocean's surface, which influence the spatial distribution of phytoplankton. To eliminate any potential ambiguity, we have made corresponding modifications in the manuscript at the relevant sections.

Thank you very much for your suggestions. It is important for us to receive these feedbacks to further improve the data set. The new version of the dataset has been re-uploaded. Please refer to https://doi.org/10.5281/zenodo.10011908. This version of the dataset consists of one NetCDF file per year. Information regarding data format, latitude and longitude, handling of missing values, units, and other data specifications has been incorporated into the netCDF files. Subsequent updates to the dataset DOI will be provided in the revised manuscript. Thank you once again for your valuable suggestions!

We really appreciate the overall evaluation, insightful comments, and recommendation by this reviewer. Our point-by-point responses to the reviewer’s comments can be found in the attached PDF file.