This paper produced a global wheat yield dataset named GlobalWheatYield4km using deep learning method (LSTM). More efforts are needed to improve the manuscripts.

1. Add a table in Section 2.2 to list all the datasets used as inputs and outputs, including the name, spatial and temporal resolutions, time period covered, the purposes of those datasets used in the paper.
2. Add a figure in Section 2.3 to describe the detailed workflow of how the dataset was produced from inputs to outputs.
3. Why the time period of the produced yield dataset is 1982-2020 instead of 1981-2021 since the AVHRR data used in this manuscript is from 1981 to 2021?
4. In this paper, only the deep learning approach LSTM was used (RF is a traditional machine learning approach instead of deep learning approach), thus change “deep learning approaches” to “deep learning approach” in title.
5. Line 15, the sentence is incomplete. “to map spatial distribution of xxx”.
6. Line 19, how 82% calculated was not descripted in the main content, and at which spatial scale, for which time period?
7. Line 19-20, the comparison of GlobalWheatYield4km and SPAM at which spatial and temporal scale?
8. Line 40, the training of statistical models needs a lot of data, and heavily depend on calibration data. Why you say that less dependence on calibration data?
9. Line 73-74, the sentence is incomplete.
10. Line 88, the GFSAD1KCM only provides a combined cropland mask, how you get the spatial distributions of wheat from this dataset?
11. Line 145, how you deal with the gaps in NDVI dataset?
12. Line 149, since the cropland mask is in 4km, almost all the pixels are mixing pixels, how you deal with that?
13. Line 185, the nRMSE is 13.1 and 16.2 for LSTM and RF, are some typos here? 13.1% and 16.2%? Same for Line 263 and Figure 3.
14. Line 196, is this the out of sample performance, at which spatial scale, for which time period?
15. Line 217, for the comparison of GlobalWheatYield4km and SPAM in 2000, 2005, 2010, it may be not fair, due to the crop yield of SPAM is the nominal value for serval years (e.g., 2010 is for 2009-2011). Please check the papers about SPAM to make sure the comparison is correct.
16. Please provide the maps of GlobalWheatYield4km and SPAM performance. Since the SPAM used less data than GlobalWheatYield4km in China, maybe the major improvement of GlobalWheatYield4km compared with SPAM is in this region.
17. Please provide the uncertainty of GlobalWheatYield4km in the released dataset also the result part.

 General comments

This study leverages phenology-based mapping of wheat spatial distribution in conjunction with both ML (RF) and DL (LSTM) modelling and global gridded weather data to produce a global wheat yield dataset at 4km spatial resolution from 1982 to 2020. The manuscript is generally good, though major revisions are required before recommendation for publication.

Specific comments

• Title – Deep learning is a type of machine learning. Correct the title to read ‘…based on machine learning approaches’ because this is accurate.
• 43 – ML approaches are a form of statistical model, so cannot be an alternative to statistical models. Could phrase as ‘ML provides an innovative approach to statistical modelling and can address…’
• 45 – Kang reference MISSING, add reference to bibliography, then rephrase so that ‘statistical models’ and ‘ML models’ aren’t seen as separate things, ML is a form of statistical modelling.
• 50 – crucial point, provide slightly more detail about the studies referenced, i.e. what crops/locations was the LSTM performance better than ML?
• 53 – if you’re saying ‘it is well recognised that…’ then at least two references supporting this statement are required.
• 71 – consider reorganising these sections because having Data and Methods as sub-sections is unusual. For example, remove the joint Data and Methods section, have separate sections for each and include the study area in the Data section.
• 98 – Table S1 should include more detail about the sources of information, this is key to the paper and the authors should consider including it in the main manuscript. Additional detail should include how the yield information was collected (farmer reports of area with observed production, farmer yield reports etc) and what organisation collected it (government ministry, NGO, UN etc) because this forms the crux of the dataset
• 109 – Expand this sentence to a brief paragraph describing the overall flow of methodology of the paper, to signpost the reader so they know what sections to expect throughout the methods section. Currently the methods section jumps around a bit and is unnecessarily confusing to the reader.
• 134 – what were the optimum hyperparameter values after tuning?
• 156 – can you clarify if this is out-of-bag RMSE for the RF? If so, please state clearly and briefly explain in the text
• 165-166 – this is a clear example of why the use of ‘statistical data’ to mean ‘observed data’ is confusing throughout this paper. Please change all references to country-reported, observed yield data to ‘observed data’ and remove references to ‘statistical data’ because it is confusing to the reader when your new dataset has been generated using statistical models. Especially confusing also on lines 122-124
• 173 – give RMSE of areas as percentages of country area rather than absolute values as these aren’t relevant when comparing between countries
• 177 – clartify uncertainties o remote sewnsing prpoducts
• 185 – were there any regions in which RF outperformed LSTM? Means of 0.72 and 0.64 are not that far apart and only regions where LSTM outperformed RF are reported, please make it explicit if RF did not outperform LSTM in any regions.
• 188 – wherever you report R² values, please also report the associated RMSE or OOB RMSE
• 242 – move uncertainties section into the results section
• 251 – go into more detail about observed yield data availability limitations – how did you overcome them and what were they precisely? Consider building into a new version of Table S1

Technical corrections

• 27 – ‘climate variability, extreme weather events and global crises…’
• 29 – ‘pandemic is estimated to have added…’
• 38 – ‘In addition’ doesn’t work here, remove entirely or substitute with ‘On the other hand’ or similar
• 53 – reword in positive manner – ‘although there are a few studies…there is still significant development to be done.’ or similar.
• 58 – incorrect usage of ‘hamper’, replace with ‘limit’ or similar
• 73 – remove pluralisations of area and production
• 118 – incorrect grammar ‘when applied it in’, please correct
• 133 – samples plural
• 186 – this is the first use of nRMSE and it is not defined (I know you defined RMSE but what is nRMSE?)
• 238 – ‘regardless’ instead of ‘despite’, depluralise years and regions

The manuscript of Luo et al. describes a new database providing gridded wheat yield at the global scale for the 1982-2020 time period. The authors used agricultural census data and advanced machine learning combined with remote sensing information and other ancillary data for the construction of GlobalWheatYield4km.

Disseminating a gridded wheat yield database is promising and the usefulness of such dataset is undoubted. However, I have serious concerns with the applied methodology. Given the fact that the presented method uses census based yield data that is disaggregated by remote sensing NDVI signal, the overall robustness of the product is almost guaranteed. In other words, the dataset shows good performance against census based yield data (Fig. 4) since census data is used to train the machine learning model. In regions where yield is high the model will provide similar high yield, and in regions where it is low it will definitely provide low overall yield, so the explained variance will be high. If the underlying remote sensing information is completely noisy (which is possible; see below) the performance of the model will be still good. And this is the case when the results are right for wrong reasons.

The methodology, as it is presented in the manuscript, is very brief and not reproducible. In order to get more information I read the cited Luo et al. (2022) paper (L108) published in International Journal of Applied Earth Observations and Geoinformation (https://doi.org/10.1016/j.jag.2022.102823). Surprisingly, the Luo et al. (2022) paper is very similar to the present study that in fact can be considered as the extension of the previous work to 54 countries (in the original paper 8 countries were included). I do not see any other added value here. The machine learning model is the same, the methodology is the same, the climate data is the same, the remote sensing information is the same. For this reason the novelty of the presented manuscript is minor.

I do not question the validity of the LSTM model. I just state that the method uses problematic input data and the results (including the correlation between the yield and the climate variables that is presented in the supplement of the Luo et al. (2022) paper) are questionable.

Overall, also because of other issues detailed below, the manuscript is not qualified for publication in ESSD. As such, I would suggest a rejection.

Major issues

The temporal coverage of GLASS LAI is not clear from the manuscript. In the Luo et al. (2022) paper it is stated that it covers 2006-2012. Due to this short temporal coverage it seems that most of the time the AVHRR-based NDVI is used which is in fact not suitable for crop type identification due to known issues with geolocation and accuracy of the several AVHRR sensors onboard the NOAA satellites. Some studies explicitly mention issues with phenology detection based on AVHRR (see e.g. Atzberger et al. 2014, https://doi.org/10.3390/rs6010257). The authors state (L85-87): “In addition, the 8 d composite Global Land Surface Satellite (GLASS) Leaf Area Index (LAI) at 1-km spatial resolution and Global Food Security-support Analysis Data (GFSAD) 1 km Crop Mask product (GFSAD1KCM) were used to map spatial distributions of wheat.” For me it means that actual (annually changing) wheat area was not estimated from the AVHRR data but rather it was handled as static; but in reality it changes year by year due to crop rotation in many places worldwide. I found this approach unacceptable. There is a major global effort to map crop types using Sentinel imagery supplemented with SAR data. It uses very high resolution and multiple data streams that are needed for accurate crop type mapping. Although I appreciate the effort of the authors, I do not see any justification that they really detect wheat cropping area and wheat phenology accurately.

In Luo et al. (2022) the authors themselves state: “Thus our wheat maps have not really characterized the dynamic variability in wheat-planting areas over time partly from avoiding the uncertainties of remote sensing data.” In my understanding this exactly means that AVHRR-based crop type identification is happening that is not justified because of the above reasons.

I have problems with small (subpixel) parcel size that is typical in many regions worldwide. Did you study parcel size distribution? Using the 0.05 x 0.05 degree resolution AVHRR data the majority of the signal will be mixed by other crops/grasses/trees/shrubs etc. I do not think that it is possible to extract usable phenology profiles from that signal in a region that is characterized by small parcels. As AVHRR NDVI is noisy and problematic even for homogeneous areas, this is just an additional source of error.

I do not see the validation of the crop type detection using ground truth. Misclassification of winter wheat is an issue even using Sentinel data. For example, barley and what have very similar phenological patterns (see e.g. Harfenmeister et al., 2021, https://doi.org/10.3390/rs13245036). So even if we assume that NDVI3g is applicable to phenology detection (which is questionable; see above), the accurate identification of the crop type is not guaranteed.

Some minor comments

L12: information on crop yield will definitely not ensure food security

L20: at this point it is unclear what is SPAM

L21: will play -> might play

L26: we are not sure what is going to happen by 2050. do not use ‘will’ here

L41: I do not agree that statistical models are less dependent on calibration data

L54: check English

L63: check English

L83: what dataset is it? NDVI3g? please specify

L98: explain the terminology. how is it related with the NUTS regions used within the European Union?

L103: PET is not actual evapotranspiration but potential evapotranspiration

L106: organic carbon content

L112: this is a key point. this statement is not true. there is very large similarity between the phenology of winter wheat and some other winter crops. see general comments

L116: I would mention harvest explicitly

L122: what statistics? I guess you mean census data. see also line 124 and other occurrences in the manuscript. check carefully and be more specific

L239 and section 2.3.2: provide loss-curve for the deep learning model. providing explicit information on the network used would have added value.

L133: provide reference to the Python package

L165: see the above comment on the terminology (statistical data). see also L166

L173: what is Kha? is it kilohectares? use standard abbreviation

L175: check English

L196: see the above comment (statistics)

L201: I guess quantifying bias would be more useful

L204: are you sure that this is the reason for the issue?

L216 and onwards: does SPAM provide annual data or is it aggregated yield for multiple years?

L254-255: I do not understand the sentence about SPAM

L234-235: check English

L251: see above; you mean agricultural census data