The manuscript “CIrrMap250: Annual maps of China’s irrigated cropland from 2000 to 2020 developed through multisource data integration” applies a random forest algorithm to classify and produce a new irrigation map product (CIrrMap250) over China at 250m resolution. The authors evaluate the new maps quantitatively and qualitatively (using reference data, withdrawal data and other existing irrigation products) over the 2000-2020 period. Generally, the paper is properly structured. It is well suited for this journal. However, the manuscript and supporting document appear rushed with several inconsistencies and mistakes. Some remarks:

• Check (and re-check) all the reported details. I.e., the performance metrics and other variables in the figures, tables and elsewhere in the manuscript (and the supplementary document). Please correct all inconsistencies. More below.
• What is the definition of ‘irrigated cropland’ as used in this study? At first I was rather intrigued when the authors mentioned in the initial sections that their product gives the irrigated cropland (which I interpreted as the fraction of vegetation cover that is actually irrigated). On further reading, however, it seemed the authors were only labeling the pixels as either irrigated [1] or not [0] and then presenting the total fraction vegetation cover (FVC) of [1] as the ‘irrigated cropland’ …is my understanding correct? If this is the case, what differentiates this product from a binary [1,0] irrigation map that is combined with the many (readily available) FVC products. Actually, one would argue that the latter method is better as it is not prone to misinterpretation by the user. Users are likely to misinterpret the produced CIrrMap250 irrigation maps to mean the ACTUAL irrigated pixel proportion and not the total vegetation cover. Also, how do you address pixels that have possibly been assigned an FVC of ~0 (e.g. at early growth stages) but have an [actual] irrigated area/extent larger than 0?
• The CIrrMap250 product is limited to China. Have the authors considered applying a similar methodology to other regions, e.g. extend it globally? Obviously, training and test datasets from other global sites would be required, but would it be viable to apply your RF classifier/model (as-is) to other regions beyond China? What would be the limitations?

Specific comments :

L16:  “… and considered the fraction coverage of irrigated cropland (i.e., the mixed pixel issue). In this study, we addressed these important gaps …” - This is somewhat misleading as the mixed pixel issue is not addressed in this manuscript. I was expecting that the authors were referring to ‘mixed pixel’ in terms of irrigation, i.e. proportion of the fraction vegetation cover (FVC) that is irrigated or not. If not mistaken, the only consideration here is the total FVC, which is provided within most RS products anyway, and can thus be similarly combined (rather straightforwardly) with any available binary/boolean [1,0] irrigation maps. Also see your comment in L495 : “CIrrMap250 cannot differentiate irrigated and rain-fed croplands at the subpixel scales. There are many small and fragmented croplands in … with complex terrain and diverse vegetation types. CIrrMap250 should be used with caution in these regions due to the wide existence of the mixed pixels”

L17: “… named as CIrrMap250 …” – consider describing all abbreviations such as CIrr before use.

L23: “… accuracy of 0.79-0.88 for years 2000, 2010, and 2020, respectively” - only for years 2000, 2010, 2020? What about the other years in between? Is it because the evaluation data were only available for those 3 years? If so, make it a bit clear here.

L42: its’ >> its

L45-46: “While numerous land use/cover and thematic cropland products have been made available to the public, they often lack information on irrigation status …” - Why would it be important to provide land use land cover (LULC) maps with irrigation status information? Should rain/precipitation or evapotranspiration information be provided within LULC maps/products as well?

L51: “…normalized difference water index (NDWI)…” - Note that there is another index that goes by the same name but used to detect floods/open water bodies (NDWI, McFeeters (1996)) – so it could ideally be used to map areas that employ flood irrigation (rice paddies, for example).

L57: “…been applied to detected irrigate areas…”  >> …to detect irrigated areas

L74: “China is a big agricultural country with the *largest irrigated area in the world …” – any reference for this?

L85: “… in paces …” – do you mean places?

L97: “many other studies “ – which studies? Add some reference[s] here

L104-105: “CIrrMap250) have a spatial resolution of 250 meters and describe irrigated cropland distribution through fractional coverage” – what of the temporal resolution? Also, as already mentioned above, this statement is misleading as one could assume you are providing the fraction of total FVC that is under irrigation.

L113-116: “These indices were generated every 16 days with a spatial resolution of 250 meters…” – to be consistent with other descriptions in the section, provide the product number of the vegetation indices product; is it MOD13Q1?

“ …band 4 …band1” – consider adding the spectral ranges here as well

L119: “Greenness Index (GI) (Supplementary Table S1)” – in Table.S1 (supplementary document) under GI, you write the ‘formula’ as GI=NIR/green,  and ‘MODIS bands’ as ‘Bands 01, 04’. The sub-caption however reads: ‘Red: band 01’ and ‘Green: band 04’ …which is which? Please correct.

L121: “… The data for unreliable pixels were reconstructed using a straightforward nearest neighbor interpolation method…” - is this the right way to go about it? For example, for an overcast pixel (which is maybe vegetated), why would you take the remotely sensed spectral signal of the next/closest cloud-free pixel (which is maybe urban/built-up)? Meaning you may end up missing vegetated pixels under irrigation or vice versa. Why not just drop such pixels from your analysis (i.e. at that particular time)?

L157-159: “In years lacking survey data, the harmonized irrigated area was determined using Eq. 2, assuming that the relative changes in statistical irrigated area are reliable” - could you explain the rationale behind Equation (2)? How to interpret it? to me it appears that a year without survey data could end up having a lower assigned/harmonized irrigated area despite having a larger irrigated [statistical] area without land survey (A_stat^t2). For instance, if we assume: A_harm^ts=20, A_stat^ts=20, A_stat^t2=30, CA_surv^ts=40 ; then A_harm^t2 becomes min(20*(30-20)/20,40)=10?

…the harmonized value (A_harm^t2) even becomes negative if we consider A_stat^t2 to be less than A_stat^ts.

What am I missing? Please clarify.

L189: “… used in combination with the MCD43A3 albedo product” - this is a daily product. Did the authors calculate the daily PET? how did you reconcile this with the other 8/16-day products?

L211: “… were then then” >> were then

L222: “*A static irrigation suitability map *were constructed based on …, and aridity index of cropland” - was this one map or several (‘*A static’ then ‘*were’). If one, why was the temporal variation of the aridity index not considered?

L230: “… (Supplementary Table S2)” – in Table S2 (supplementary document), why do you have the same ‘Suitability value’ for the lowest suitability classes S3 and S4. I.e., for the ‘elevation’ and ‘slope’ irrigation suitability factors, S3=2 and S4=2.

L258: “and time-invariant environmental variables (i.e., latitude, longitude, crop intensity” – why is the crop intensity considered time-invariant?

L261: “To enhance the accuracy of these maps, a spatial filter (a 7x7 window)…” - clarify what you mean by this. Why 7x7? …‘constituting <5% of the window area’ is ambiguous. Is the 250 m resolution retained after this?

L276: “were acquire from …” >> were acquired

L282-: “Due to the lack of georeferencing information, we georeferenced these land use maps using the georeferencing tool in ArcGIS in conjunction with high-resolution images “ – the authors do not talk about the data that were used to serve as ground control points for the georeferencing (e.g. How many GCPs, their spatial distribution, …?)

L294-: “It’s noteworthy that this percentage represents the proportion of cropland within the 250 …, not the proportion of irrigated cropland to total cropland” ; L362: “irrigated cropland in CIrrMap250”. As already mentioned, giving the irrigated cropland as a percentage is very likely to mislead users into assuming that your irrigation product provides the proportion of the total fraction of vegetation cover (FVC/cropland) that is irrigated. If feasible, wouldn’t it be more useful to have both products, i.e. the total fraction cover product and the proportion of that that is deemed irrigated? The authors also acknowledge in L492 that “…cirrmap250 has a relatively coarse resolution”. You may still argue that at the relatively higher spatial resolution of 250m, one could assume the whole cropland (total FVC) to be equivalent to the irrigated area. This might be true but still needs validation to avoid being misleading.

L340: “… under severe to extreme…” - In the previous sentence (L339), only low, moderate, high and severe WSI ranges are described. What is the extreme WSI range? Is extreme synonymous to severe here?

L352: “CIrrMap250 and IrriMap_CN performs similarly in user’s accuracy…” – TableS5 (supplementary document) shows  a user accuracy (UA) of 1 (error of commission=0). Can this perfect UA be explained? From Fig3c/TableS2 (year 2020), IrriMap_CN has a producer accuracy (PA) of 0.2, why this huge discrepancy between the [perfect] irrigated.UA (1) and the [rather poor] irrigated.PA (0.2)?

Since you use ~20,000 samples in your classification exercise (into irrigated and non-irrigated), could you provide (in supplementary doc) the CIrrMap/IrriMap confusion matrices for 2000, 2010, 2020 to aid with interpretation (i.e. how many of the reference samples are irrigated or not? How do you split these into training and test sets? …more details on how the RF classifier used in CIrrMap250 performs, …)

L366-367: “CIrrMap250 yields irrigation ratios (i.e., the ratio of irrigated area to the total cropland area) of…” – this sentence contradicts L294 (i.e., “… this percentage represents the proportion of cropland within the 250 …, not the proportion of irrigated cropland to total cropland”), and many other statements in this report (e.g. L495 “cirrmap250 cannot differentiate irrigated and rainfed croplads at the subpixel scales”). Such inconsistencies make it somewhat difficult to follow and interpret your results/analyses.

L370-374: “However, CirrMap250 tends … southern part of South China (SC)” – why? Could you discuss this section a little bit more. Readers may not go back to the literature on the other products to find out by themselves. Also, what does ‘*southern part of *South China’ mean?

L388: Fig5 – This figure needs improvement. How come no irrigated pixels in zone B are detected by the 1Km GFSAD product?

L391: “Figure 6 …CIrrMap250 exhibits a robust agreement with OPTRAM3” - This is not clear from the figure. Qualitatively, Figure 6a may even be interpreted differently unless the authors have overlain CIrrMap250 over OPTRAM30. If that is the case, please find a better way to illustrate/present the map inter-comparisons.

Figure6 – the [0-100] color scale as provided in Figure4 is missing. In supplementary, Figure S2 (b, c, d)   – the magenta color scale (for IrriMap, IAAA, GFSAD) is missing.

Additionally, why was year 2019 selected in Figure6a,b (CIrrMap250/IrriMap_CN) for the comparisons with the 2014-2020 OPTRAM30 product?

L406-407:  “…, namely 2010 and 2020. The estimates of irrigated areas from the other two maps, namely IAAA and GFSAD, are able to explain only a small proportion of the variances in irrigation water withdrawals (i.e., 0.12 and 0.20) …” – Please clarify. According to Figure 7c,e, these (IAAS and GFSAD) metrics only apply to year 2010 NOT 2020.

L410: “…irrigated area estimates against irrigation water withdrawals…” – maybe you mean ‘irrigated water withdrawals against irrigation area estimates…’? Y against X.

L426: “As shown in Figure 9, all subregions exhibit an increasing trend in irrigated area from 2000 to 2020” - is this conclusion based on CIrrMap250 or some other [reference] data?

L435-: Figure 9d - some of the percentage entries in the concentric pie charts are likely incorrect. I.e. percentages for years 2000 and 2020 add up to 101% (11+7+17+30+24+12) and 98% (11+11+16+26+22+12), respectively.

L445: “… The net expansion of irrigated area is about 180,000 …” but L427 reads “The irrigated area of China increases from 750,000 to 950,000…”, which is ~200,000. Both for the 2000-2020 period. Please be consistent with the presented numbers.

L465: “…leading to a decrease in irrigation mapping accuracy by 8%-26% (Supplementary Figure S4).” – do these numbers refer to supplementary Figures S3? They do not appear in Figure S4.

The caption of Figure S3 reads “Comparison of irrigated ratio estimates of CIrrMap250 and IrrMap_CN in China, Northern China, Xinjiang Uygur Autonomous Region” … does this mean that this conclusion only applies to that specific part of China?

L474-475: “... The accuracy of the irrigated cropland map would decrease by approximately 5%-6% (Supplementary Figure S6) “ - Figure S6 (in the supplementary document) contradicts this statement. From the figure, it appears that “considering FC of cropland” (blue bars according to the plot legend, and “this study” according to the caption) yields worse overall accuracies (OA) than “Neglecting FC of cropland” (green bars). This is the case for all three (2000, 2010, 2020) years. Is the plot legend correct?

Reference

McFeeters, S. K. (1996). The use of the Normalized Difference Water Index (NDWI) in the delineation of open water features. International Journal of Remote Sensing, 17(7), 1425–1432. https://doi.org/10.1080/01431169608948714

Please see my comments attached in the .pdf file.