The manuscript presents a dataset useful to train machine learning models for LULC mapping and the methods and results it presents are original and sufficiently well explained. However the manuscript would benefit from a clearer focus on the scope and limitations of this training dataset. For instance, both the title and the abstract do not contain ‘training’ and do not obviously point to the type of data that are presented. Some of the technical choices should be better justified (e.g. why not using the standard Land Cover Classification System as the basis to derive consensus between different land cover and legends; why not including the LCCS-based types from the MODIS land cover; some of the LULC categories are not defined, e.g. broadleaf cropland). The manuscript should more openly discuss the limitations and weaknesses of the dataset including on the methods applied for data validation and on the potential applications of the dataset – for instance it is not immediately obvious if the dataset could support machine learning methods to detect LULC changes with complete transition matrixes. Particularly for the classes with lowest purity – defined as the combined consensus spatially and temporally across products, if would be useful to add to the discussion some insights on what classes cause confusion and reduced consensus. I believe this (ancillary information) may be included as part of classification efforts. This information could be enriched by I recommend the publication of the manuscript after major revision – see my specific comments and suggestions in the attachment. 

Review of TimeSpec4LULC: A Global Deep Learning-driven Dataset of MODIS Terra-Aqua Multi-Spectral Time Series for LULC Mapping and Change Detection

Rohaifa Khaldi et al.

General comments

The manuscript is generally well written and organized. I recommend it for publication if the following are addressed. This does not require further processing and/or analysis, just clarification.

There are no reference to MODIS LC product at all nor of the ATBD:

Strahler, A., D. Muchoney, J. Borak, G. Feng, M. Friedl, S. Gopal, J. Hodges, E. Lambin, D. McIver, A. Moody, C. Schaaf, and C. Woodcock. 1999. MODIS Land Cover Product Algorithm Theoretical Basis Document (ATBD) Version 5.0. Boston: Boston University, 89 pp.

Per validation / accuracy assessment: Expert opinion is extremely problematic. There is no mention as to cross-validation, agreement of interpreters etc. Is this just a call as to what class it might be?

Muchoney, D.M., A. Strahler, J. Hodges and J. Locastro. 1999. The IGBP DISCover Confidence Sites and the System for Terrestrial Ecosystem Parameterization: Tools for Validating Global Land Cover Data. Photogrammetric Engineering and Remote Sensing 65 (9): 1061-1067.

Co-registration: There is no documentation of pixel-to-pixel co-registration. Did you look at the PSF?

There are no problems reported: Might there not be mention of assumptions, possible errors?

There is no details on the classification algorithms.

Monthly composites: Are not 32-day monthly MODIS composites available?

Specific comments

10        I don’t know what “smartly pre-processed” means

25        Not sure what an “essential planetary boundary” might be

45        Gglobal products are not meant for local studies. MODIS was meant to parameterize  BGC and GCM.

60        Deep Learning is not defineds.

80        CNN and RNN are not described. NNs have been in use for quite some time, at least 20 years

90        The purpose overall is to create a dataset that allows for deep learning?

125      Cross-walking categorical/nominal variables, i.e. classes is deeply problematic

185      What about areas with persistent cloud cover: Ecuador, Colombia, Peru, DRC, ROC?

200      Again, global products are not necessarily meant to be down-scaled

Technical corrections

• Graphics: While this may just be the pdf rendering, the graphics are small and color rendering dull: Figures: 2, 3, 4 5