L47-48: To our knowledge, this is the first set of published 30 m annual land cover and land cover change datasets that span from the 1980s to the present for the United States. This sentence should be revised because a recently paper was not referred in this manuscript. Prof. Gong’s team produced the first 30 m global annual to seasonal land cover maps for 1985–2020, which covered the study area and temporal extent of LCMAP.

Reply: Agreed with the comment about the global land cover. The Global products produced by Lui et al. (2021) and Zhang et al. (2021) included seamless data cube and land cover for the globe. However, the LCMAP products contain annual land cover and associated spectral change information that has a wide implication for land change assessment. We changed the sentence and added and relevant manuscripts in the reference list. Thanks for the recommendation.

L293: 3.4 Land cover classification. It is better to add a land-cover classification flowchart and the corresponding explanations here.

Reply: A flowchart has been created and added as Fig.3

L329-333: How many bands and features were used in total? Why the ‘intercept’ parameter was discarded? What is the contribution of brightness temperature bands?

Reply: There was a total of 68 feature combinations for classification. We used model information generated from 7 spectral bands, including 8 model coefficients and one RMSE for each band. We also used five ancillary data: elevation, slope, aspect, position index, and potential wetland index. 

The classification was accomplished using information from CCD models. These models were processed to make the land cover classification because different land cover classes can have different shapes for the estimated time series models, in which the overall mean and model coefficients except intercepts can be used to estimate the intra-annual changes caused by phenology and sun angle differences for the ith Landsat Band. The classification was accomplished without using intercepts. 

Because each tile has a different classification model, the contribution of the brightness temperature band may vary depending on the land cover type, condition, time, and location.

L366-368: QA/QC is extremely important for the validation dataset. How many interpreters are assigned for each sample? More details should be given for the QA/QC processing, not just a reference (Pengra et al., 2020a).

Reply: The 25,000 reference sample pixels were each interpreted by a trained interpreter with approximately 60% of these pixels interpreted independently by a second analyst.  Much of the QA/ QC process relied on comparing the interpretations at these duplicated sample pixels. Duplicated sample pixels that had interpreter disagreement were used in QA/QC process such as: to identify issues with specific classes or interpreters, to flag sample pixels for further review and possible editing, and to provide ongoing training and feedback to interpreters throughout the collection process.  QA/QC related reviews were also completed on sample pixels that showed interpretation data such as uncommon and/or illogical land use and land cover combinations, multi-year disturbance processes, rare classes, or other opportunistically identified situations.  Sample pixel interpreted attributes were edited if necessary to create the final attribute assignments for the reference data.  These final attributes were then cross-walked to a single LCMAP land cover class label. We added several sentences to explain QA/QC process.

The number of interpreters total ranged between 5 and 11 depending on the specific point in time of the collection effort.  QA/QC reviews and edits involved between 2 and 4 interpreters depending on the type of review and point in time in the collection effort (based on interpreter availability and other factors).

L524-526: Indeed, the mapping error of NLCD could potentially be carried over to the training samples. How does it affect the classification accuracy of the LCMAP dataset? Two references can benefit this concern (Fig.1 in Gong et al., 2019; Fig. 10 in Zhang et al., 2021).

Reply: When the method was implemented to produce LCMAP products, we selected NLCD as the training data source. The reference dataset was developed independently and was not completed when LCMAP products were produced. To ensure the classification can capture land cover features correctly, a certain number of training samples were selected in each of 3x3 tiles. The large number of selected samples that were widely distributed in the target tile represent land cover conditions of each land cover type. We did not use reference data to check these training samples due to objective concern for reference data. However, we appreciate the reviewer’s suggestion for these two references. We also added several paragraphs in the discussion session to discuss training data selection and cited the two publications.

“To select a sufficient size of training samples is important for CCDC models to obtain accurate classification. Previous land cover post-classification analysis suggested that the overall classification accuracy increased when the training samples increased (Gong et al., 2020). The recent global land cover classification also suggested that the appropriate training sample size for a mapping extent of three 158 km x 158 km tiles should be larger than 60,000 (Zhang et al., 2021). For the LCMAP land cover classification, a much large size training was utilized to ensure that these training could represent landscape features in the classification tiles.”   

L488-489 The LCMAP product suite includes five land cover change and five land surface change science products. There is no link to the corresponding parts in the Supplementary Material. Please check throughout to link the supplementary file with the main manuscript.

Reply: The Supplementary Material was attached to the manuscript when it was submitted to the journal. We have checked with the topic editor and was told that the attachment was appropriate. We had words to explain the supplementary material is attached.

General comments:

Xian et al. present the new Landsat-based LCMAP annual 30-meter land cover and change dataset for the conterminous United States, spanning 1985-2017 (which on the LCMAP website now seems to go through 2019). Based on the state-of-the-art Continuous Change Detection and Classification (CCDC) algorithm, the product represents a large step forward in semi-operational, large-scale monitoring of land cover and its change through time. The data are well described and easily accessible, with well-defined accuracy and uncertainties (both in the manuscript itself and in the data layers and metadata). This clearly represents an important and broadly applicable resource for a wide range of applications. I have a few minor suggestions for further clarification and on the presentation in the manuscript, but overall, this is an excellent contribution that will almost certainly become a very important dataset for both science and management.

Reply: Appreciate your comments.

Specific comments:

Lines 103-105: Are there specific examples that the authors could provide here?

Reply: We added several references to enhance the continuous monitoring approach. Thanks for the suggestion.

Lines 153-155: The Introduction ends on a bit of a weak note. I’d suggest closing with a clearer statement of the objectives of the manuscript and the bigger-picture importance of the dataset, particularly following up on how the “Lessons learned” (Brown et al. 2020) informed the implementation of the LCMAP data presented here in this manuscript.

Reply: Agree. These sentences were changed by including “Lessons learned” from prototype development that was presented in a previous publication. 

Lines 312-314 and 519-521: Why were grass and shrub included in a single grass/shrub category rather than separated into two classes? It seems like grass and shrub would likely be both ecologically and spectrally distinct, so some discussion of why that decision was made would be helpful.

Reply: We added several sentences in 312-314 to explain the reason of combing grass and shrub together with consideration of challenges in spectral feature in the western US. Also, in the discussion section (519-521), we explained uncertainties of NLCD 2001 grass and shrub especially in the western US. The LCMAP product has achieved consistent and relatively higher accuracy by combining grass and shrub as a single class in the classification. 

Lines 472-473 and Fig. 9: Could the authors discuss or speculate about why the overall accuracy seems to decrease monotonically through time (albeit, quite a small decrease) starting around 1997? They address why the accuracy in 2017 decreases quite suddenly (limited Landsat observations at the end of the time series), but it’s not apparent to me why there would be a long-term monotonic decrease prior to that rather than exhibiting more-or-less random variation from year-to-year.

Reply: It is complicated for causes of the overall accuracy change. We believed that several factors could influence the overall agreement between reference data and mapping result. Factors including the CCDC models, Landsat data quantity and quality, training data used, the availability of high-resolution image used by interpreters to generate the reference data, and specific land cover type could impact the overall validation accuracy. We added several sentences to explain potential reasons caused the overall accuracy.

 Fig. 5b, 7d-f, and 8d-f: I find it very difficult to interpret these figures. The mix of gray-scale confidence shading is tough to distinguish from the different colors, and it is nearly impossible to distinguish the two shades of purple in the maps for increasing and decreasing vegetation. Those two colors are so similar to each other that it is extremely difficult to tell them apart in the maps.

Reply: We modified all graphics of land cover confidence by changing the dark purple color to green to make the two types of confidence more distinct.

Fig. 5d: I would suggest rephrasing the caption to read “…(d) total number of _spectral_ changes detected…”

Reply: Changed the words as suggested.

Fig. 7g-h: I would suggest making the caption more descriptive about what these represent. The number of changes through time? And are these the number of thematic changes (e.g., like Fig. 5c) or spectral changes (e.g., like Fig. 5d)?

Reply: Captions of both original Figs 7 and 8 are changed to represent the change information.

Technical corrections:

Lines 63-64: I would suggest removing this first sentence of the Introduction. The second sentence is a much stronger opening, in my opinion.

Reply: Moved the first sentence to the beginning of the second paragraph.

Line 406: I don’t see any dark green in Fig. 5b. (see also comment on 5b above in the specific comments)

Reply: The graphics were changed to show green color.

Line 416: I would suggest adding a reference to Fig. 5d after “…in the east”.

Reply: Agree. Added a reference of NLCD land cover change here.

Line 430: I would suggest adding “(respectively)” after “2008 and 1995” to make it clearer that the increasing trend ended in 2008 for the grass/shrub and 1995 for the tree classes.

Reply: Agee. Added the suggested word.