General comments

The authors present an interesting effort to improve current estimations of burned areas in croplands, which are not well mapped by global BA algorithms, particularly by coarse resolution sensors. They use a similar approach to GFED4s/5s products, which estimate burned area from active fire information using a set of calibration sites. The study strongly relies on appropriate maps of croplands, which the authors did not produce, neither validate. This is an important gap of the manuscript, since using a coarse resolution global land cover map that includes also errors adds a significant source of uncertainty, as the authors acknowledged, but not quantify. This should be include in the new version of the manuscript.

The authors base their methodology in MODIS AF, but it is well known that this sensor is close to its functional life. Therefore, I strongly recommend the authors to use VIIRS data instead, or even better from a combination of the two to assure both time series length and future extension.  Otherwise, the GloCAB product would just be historical estimation. I realize the convenience of having a long temporal series of crop fires, but this should be balanced with the reduced accuracy from using a coarse resolution AF product.

Related to this, the methods do not clearly describe how the AF/BA ratio was applied to just the cropped areas, as within a 0.25º cell many wildfires may occur simultaneously to agricultural burnings. Did you use the cropland area defined by the MOD12 product to mask only the cropped area? What would happen then with the cropped areas where small parcels (< 25 has) are widely extended?

In addition, it is not clear either how did the authors extent the conversion factors to other climate regions. Did they apply the winter wheat coefficient to all winter wheat worldwide, regardless the continent-climate zone where they are located?

I appreciate the difficulty of validating the product. However, the exercise that is included in the paper is clearly insufficient to grant any significant confidence to the results. I suggest the authors to generate a few additional agricultural BA maps, similarly to those used for calibration, or at least compare iteratively their results with their calibration sites with a bootstrapping approach.

Specific comments

Line 48. Perhaps the authors should quote the latest version of GFED.

Line 68. With the same logic of quoting the algorithm description for the MCD64 product, the authors should quote Lizundia et al., paper for the FireCCI51.

Line 120. Include references of this statement, which is very important as it justifies the selection of the crop types being considered.

Section 2.4 seems more appropriate for the methods section, as it was part of your own developments for the product.

Line 191: “training reference areas, several challenges limited the mapping. Most importantly, small fields and poor air quality in several countries (e.g. India and Thailand) prevented the analysts from observing changes in the fields even with 3-m Planet”. This is a bit confusing, since you did not include any calibration site in India or Thailand, according to table 1.

References

Lizundia-Loiola, Joshua, et al. (2020), 'A spatio-temporal active-fire clustering approach for global burned area mapping at 250 m from MODIS data', Remote Sensing of Environment, 236, 111493.

General comments            

GloCAB is an interesting and novel approach to generate the first global product dedicated to spatial estimates of the cropland burned area. It will be a useful complement to the forthcoming release of the GFED5 (preprint also available in the journal) even if, as the same authors have discussed, this is a preliminary attempt and additional work is needed to improve and refine the product. Below are more detailed comments and points that I think should be addressed before publication.

Emissions from cropland fires. At this stage, the product is not mature enough to generate accurate estimates of GHG emissions from the burning of crop residues. Robust and global Tier I estimates of emissions for the same crops in the paper are already available (e.g. FAOSTAT, EDGAR) based on the area harvested reported from countries. It is recognized that these emissions represent overall a small proportion of agricultural emissions. Yet, the authors decided to open the abstract with a sentence on emissions (from general fires). I would recommend to change the focus in the narrative to other and more important aspects which the product can already address and that the authors already described in the introduction (e.g. use of this information in monitoring systems; health implications; the applications for more sustainable agricultural practices). 

Land Cover. The authors used the IGBP land cover type from the MODIS land cover, collection 6. However, the 3 LCCS land cover layers, as reported by Sulla-Menashe and Friedl (2018), are instead the reference type of the MODIS land cover product and those with the higher accuracy. The authors should specify why they decided to use the IGBP type instead. The authors treated the mixed class 14 as full cropland. I tend to agree with this approach given the nature of the data, but it may be worth specifying the reasons for this choice. While I wonder if the authors explored other alternatives (e.g. ESA CCI), I agree that discrete global land cover classifications are all likely to suffer from omission and commission errors and importantly, uneven performances across regions (which the authors should also discuss). Possibly, dynamic land cover products that are better aligned temporally to fire dynamics might be something to look at in future developments. Reference for Sulla-Menashe and Friedl is missing from the bibliography. 

Cropland fires. The authors have included some information on the characteristics of cropland fires. However, I believe it would be beneficial to present this in a dedicated section discussing the main differences between pre-planting, pre-harvesting and post-harvesting fires; associations with crop type (e.g. pre-harvesting fires in sugar cane fields) and agricultural practices as well as the implications for detection (e.g. pre-planting fires are followed rather rapidly by soil preparation which alters the burned area and reduce the ability to detect the area that was burnt). Information on the prevalence of geographical distribution by crop type and gaps in literature would be also useful.

Trends and Africa. The abstract does not discuss the role of Africa in the assessment of global trends and the manuscript presents global results with and without Africa only in the supplementary section. This region certainly contributes significantly to the uncertainty of this product, including due to the prevalent small size of crop fields in the region. A more structured discussion of the caveats of their approach in Africa and the impact of this region for the global estimates would be beneficial for the transparency and readability of the manuscript. 

• Major comments

The GloCAB dataset offers the first global cropland-focused burned area product, providing monthly data from 2002 to 2020 at 0.25° resolution, targeting small-fire types in croplands. This first-of-its-kind effort offers a detailed view of small-fire types in croplands, highlighting region-specific trends and paving the way for improved understanding and policy-making. I would like to suggest a few more things that need to improve in this study.

The author can highlight the novelty of cropland-focused burned area mapping further. In the current version, it is difficult to see a significant difference from previous products. It would be nice to have more discussion about the implications that only GloCAB has. Also, as the land cover product influences the overall results, the author quantifies the uncertainty of cropland-burned area mapping using different land cover datasets.

• Minor comments

Line 18-20: I agree with the author’s point about cropland-focused burned area mapping. But this sentence does not link smoothly with the previous context

Line 73-75: As MODIS will be decommissioned, VIIRS will be used alternatively. For long-term analysis, intercalibrating between MODIS and VIIRS is needed. That information could be useful for readers.

Method: I think the reliability of GloCAB heavily relies on the performance of MODIS products. The author can simply add quantified uncertainty of each input dataset.

Figure 7, 8, and 9 should be improved with clear color and formal legend. Statistical values also can be included in the figure.