L233. Correspondence --> Confusion matrix. Correspondence is not the standard term for this table. Also font used is different. Change to Times new roman for the table

Response: We thank the reviewer for their comments. Correspondence matrix has been changed to Confusion matrix. Text has been changed to Times New Roman.

L242.Correspondence --> Confusion matrix. Correspondence is not the standard term for this table.

Response: Correspondence matrix has been changed to Confusion matrix.

1. In the manuscript, you stated that “LCM2021 methods replicate those for LCM2017 to LCM2020 with minor deviations to enhance accuracy.”. According to the LCM2020 product description (https://www.ceh.ac.uk/sites/default/files/2021-11/lcm2020ProductDocumentation.pdf), the overall accuracy improved from 79.2% (LCM2020) to 86.5%. Apparently, some changes in the processing pipeline have been included (e.g., the purity across precedent years was set to 99%, instead of 80%). But, the model (Random Forest) has been used in both cases. Please, highlight the changes made and try to relate those changes to the performance improvement.

Response: We thank the reviewer for their comments. Increases in classification accuracy have been made between LCM2020 and LCM2021, however the figures in the reviewer’s comments are not directly comparable; the 79.2% accuracy figure for LCM2020 refers to the 21-class nomenclature, whereas the 86.5% accuracy figure for LCM2021 refers to the 10-class aggregate nomenclature. The accuracy for the LCM2021 21-class nomenclature is 82.6%, so whilst the classification accuracy for LCM2021 has been improved in comparison to LCM2020, the directly comparable level of increase is 3.4%, rather than 7.3%. We think these accuracy gains are due to improvements in the cloud-masking and training data distribution, particularly the increase in training areas in areas where the training data may have been too sparse previously, however we do not have quantitative evidence for this. The remit of this article is the LCM2021 data set and its production, not previous products, however the comment “LCM2021 methods replicate those for LCM2017 to LCM2020 with minor deviations to enhance accuracy.” has been updated to “LCM2021 methods replicate those for LCM2017 to LCM2020, with minor deviations in cloud-masking processes and training data sourcing to enhance accuracy” to highlight these changes.

1. You have trained a Random Forest to classify the land cover. However, state-of-the-art research papers often use deep learning-based approaches to do this task, showing a significant performance improvement [1] [2].

Response: This is an interesting suggestion. The strength of convolutional networks (as applied in [1,2]) is primarily in understanding the structure of image patches. Urban areas having a high structural component, so such methods work well. However, many of the land cover classes we map are less structured and the information content is primarily in the spectral and temporal elements of the satellite data; Random Forest is highly suited and well-established for such tasks. However, work such as that by Pelletier et al., (2019) and Zhang et al., (2023), suggest that 1-D convolutional networks may be more appropriate for our requirements, and we plan to test them in the future.

References:

[1] Pelletier, C., Webb, G.I. and Petitjean, F., 2019. Temporal convolutional neural network for the classification of satellite image time series. Remote Sensing, 11(5), p.523.

[2] Zhang, H.K., Roy, D.P. and Luo, D., 2023. Demonstration of large area land cover classification with a one dimensional convolutional neural network applied to single pixel temporal metric percentiles. Remote Sensing of Environment, 295, p.113653.

1. The data set is imbalanced. Thus, please, include kappa to evaluate the accuracy performance.

Response: Kappa has been added to Tables 3 and 4.

And some minor comments:

1. It would be nice to have not only the majority, but also the top 5 predicted classes (with the respective probabilities), as this could help further evaluating machine learning models (e.g., use this data as a benchmark) and also to evaluate some other downstream machine learning tasks.

Response: We thank the reviewer for their thoughts on our data and its potential reuse. In previous LCM versions (LCM2000 and LCM2007) the probability of the top 5 classes was included, however this information was scarcely used and so this aspect of the data set was dropped when LCM production moved to using Random Forests. Therefore the top five predicted classes and their probabilities are not given in LCM2021, however if there is demand from the user community for data of this nature, then we can look to include this in future LCMs.

1. Would it be possible to colorize the confusion matrices (Table 3 and 4)?

Response: We thank the reviewer for their suggestion and previous examples, however on this occasion we would prefer to keep the confusion matrix not colourized. We acknowledge that there are a variety of ways in which confusion matrices can be presented, however here we wish to preserve the absolute numbers of validation points within the confusion matrices, rather than normalising the values to a scale that would work better for colourisation, or presenting confusion matrices containing only producers accuracy as in the referee suggested reference Liu and Shi (2020). For the ranges in absolute values contained within the confusion matrix (0 to 10,102 for Table 4), colourisation would result in the majority of table cells comprising low values being presented as a similar colour, so not being particularly informative. Presenting confusion matrices un-colourized is also consistent with how confusion matrices have been presented in previous ESSD publications such as Zhang et al. (2021) (Tables 3-5), Li et al. (2023) (Table 3), and Liu et al. (2023) (Table 3).

References:

[3] Zhang, X., Liu, L., Chen, X., Gao, Y., Xie, S., and Mi, J.: GLC_FCS30: global land-cover product with fine classification system at 30 m using time-series Landsat imagery, Earth Syst. Sci. Data, 13, 2753–2776, https://doi.org/10.5194/essd-13-2753-2021, 2021.

[4] Li, B., Xu, X., Liu, X., Shi, Q., Zhuang, H., Cai, Y., and He, D.: An improved global land cover mapping in 2015 with 30 m resolution (GLC-2015) based on a multisource product-fusion approach, Earth Syst. Sci. Data, 15, 2347–2373, https://doi.org/10.5194/essd-15-2347-2023, 2023.

Liu, C., Xu, X., Feng, X., Cheng, X., Liu, C., and Huang, H.: CALC-2020: a new baseline land cover map at 10 m resolution for the circumpolar Arctic, Earth Syst. Sci. Data, 15, 133–153, https://doi.org/10.5194/essd-15-133-2023, 2023.