Summary: This study presents a new method for estimating latent heating (LH) profiles from geostationary radiances, and compares the result with established methods that use NEXRAD ground-based radar and TRMM and GPM spaceborne radar. The methodology for estimating LH is similar to what is used for TRMM/GPM LH profiles and is based on a database of output from a convection-permitting resolution model. The authors find that the GOES-based LH estimates are similar to those obtained from NEXRAD and GPM, and produce similar (positive) impact on model forecasts when used in model initialization.

General comments: The use of geostationary data to estimate latent heating is interesting, and potentially valuable, as the Geo data provides much more extensive spatial and temporal coverage relative to NEXRAD and GPM. I think this manuscript is publishable, but needs to be supported with quite a bit more explanation of the tools and datasets used, and also should contain additional context and caveats. I would like the authors to consider the following general recommendations.

1. Any LH estimate from remote sensing is by nature indirect - the observation is of the result of a process that involved LH, not of the LH itself. For example, there can only be hydrometeors for the radar to observe after the condensation process has already happened. A change in the cloud top brightness temperature can only happen after the air has arrived at the top of the storm (having already gone through the condensation process). Please comment on this - I think there is a significant unanswered question that relates to the time and space disconnect between an observation of the result of LH and the LH itself. 

2. Convection is identified using time sequences of GOES imagery, yet the LH profiles are binned by the magnitude of the cloud top brightness temperature. It seems to me that an interesting and more direct comparison could have been made between the simulated LH and the simulated time-difference brightness temperature. Please discuss.

3. An obvious point of concern in any model-based lookup table is the model construction and configuration. A 3 km horizontal grid spacing is barely convection permitting, and most simulations of deep convection meant for scientific analysis are now conducted at grid spacings of 1 km or less (most often smaller than 250 meters). Studies comparing simulations with sub-1-km grid spacing with those run at ~3 km have consistently shown that updrafts in 3 km grid spacing simulations are too wide and often too strong, and that the latent heating distribution is shifted higher in the coarse resolution runs relative to the fine resolution runs. In addition, studies have also shown that the LH position and magnitude are very sensitive to the details of the cloud microphysical parameterization. I have a number of questions that I would like the authors to address:

- Why did you not run the WRF model at finer grid spacing? Even if this was not computationally feasible, at least one simulation should be run at fine grid spacing and the LH characteristics compared to assess sensitivity.

- What was the sensitivity of the simulated LH to choice of microphysics? I do not expect a detailed study of this, but as with the previous question, one could imagine running companion simulations of the same case, one with Thompson microphysics and another with (for example) Morrison or WSM6. This would at least provide a first order estimate of the sensitivity.

4. There was not enough detail provided about the simulation database itself. I was missing the following, which the authors should provide:

- What were the lengths of the simulations (in time)?

- What were the geographic domains?

- Which data was used for initial and boundary conditions?

- What was the model output time frequency?

- How many vertical layers were used? (this can have as large or larger effect on the convection than the horizontal grid spacing)

- How were the simulations validated? How did the authors ensure they provided a reasonably realistic depiction of storm structure?

- Did the simulations span a range of convective types (size, longevity, mode of organization)?

5. The various LH estimates seem to reflect different sources of information on LH. For example, NEXRAD is sensitive to large hydrometeors and primarily obtains information from the lower portions of the troposphere. As such, one would expect the NEXRAD estimates to be biased toward the lower portion of the storm and miss LH in the middle and upper portions. TRMM/GPM radars operate at a shorter wavelength - they will see more of the smaller hydrometeors higher in the storm and may miss some of the heaviest rainfall due to attenuation). One would thusu expect their information to come from the middle portions of the storm but perhaps miss the very lowest and highest layers due to missing detection of heavy rain and small cloud particles. Geostationary data only sees the change in cloud top properties - it's not clear which portion of the storm produces the change at cloud top, but it is likely weighted toward the middle and upper portion of the storm. I would like to see the authors comment on this, and to perhaps discuss how the three sources might be merged in those instances where all three view the same place and time.

6. There were no caveats listed in the conclusions - one would expect that there are places and times where the GOES data might provide a more reliable estimate of LH and others where these estimates will have larger errors. What are these? Also, there was no mention of future work - what is next? This should also be discussed in the conclusions section.

Specific comments:

1. June 2017 (the case used to assess impact) is within the time frame used to run the WRF simulations that form the database of profiles. In testing a database-based method, it is common to test on a case that lies outside of the training dataset. I wonder what the results would look like if you compared the estimates for a month from 2019?

2. It was clear that there are discrepancies between the NEXRAD and GOES detections of convection. It would be interesting to see statistics on how often these discrepancies occurred.

3. The scatter in the plot comparing GOES vs NEXRAD LH in Fig 8 is very large. It is surprising that the correlation was ~0.8. I wonder if the relationship is more obust for smaller LH values than for larger? I suggest using log-log axes for Fig 8 to better be able to examine the smaller LH values.

4. The phrasing in lines 560-560 on page 16 is confusing - it makes it sound like you are replacing the observed LH with the LH from the model. I think that what you are doing is inserting the observed LH into the model (replacing the modeled LH), right?

5. Follow-up question - are you inserting the observation-estimated LH profile? If so, the NEXRAD profile would be bottom heavy while the GOES profile would be top-heavy, right? This would explain the precipitation differences, I would think... NEXRAD LH would produce warming lower in the troposphere, which should result in a much larger effect on buoyancy, relative to GOES.

6. While the magnitudes are similar between NEXRAD and GOES estimates of LH, the position of the peak in the vertical matters quite a bit for large scale dynamics. How has this discrepancy been addressed in the literature? Is it assumed that NEXRAD is biased low? Is CSH (and by extension GOES) biased high?

Review for AMT of Lee et al. manuscript entitled “Latent heating profiles from GOES-16 and its impacts on precipitation forecasts.”

General Comments:

Latent heating (LH) is an important process-level cloud variable.  To address the temporal gap in LH observations (arising due to satellite estimates only being available periodically), the authors develop a new GOES-based LH retrieval.  They compare their GOES LH retrievals with existing satellite and NEXRAD estimates, and then demonstrate the impact (on precipitation forecasts) of assimilating the new LH into WRF.  The analyses of impact on precipitation forecasts are fairly minimal (skill scores are shown for a few boxes for a few forecast hours), so this part could be thought of as a proof of concept for using GOES LH in WRF. 

My major questions are:

1. Why is only 1 month of data compared across the products? Can additional months be aggregated with some averaging over connected convective features to remove the substantial noise that shows up in instantaneous pixel comparison plots (e.g., Fig. 8)? 
2. There is a lot of discussion of this quantity: total LH. As best I can tell, this is not an average convective LH, nor is it a vertical integral.  It is a sum of convective LH profiles in a box.  I did not understand the purpose of just showing the sum, which is largely a result of summing profiles over different convective area counts (and the convective area counts are product dependent, being different for NEXRAD, CSH and GOES).  The average profile structure and convective area differences should be portrayed and analyzed separately.  If my understanding of total LH computation is incorrect, then it needs to be clarified through the manuscript. 
3. The analysis of the impact on precipitation forecasts is pretty minimal, and as a reader, I was not convinced that the abstract text stating “improving the forecast significantly” is warranted quite yet. I would re-phrase the text in parts to suggest that this is more of a proof of concept or demonstration of the potential value of assimilating LH.    

I did not comment on grammar, but I emphasize that a comprehensive read-through is needed to correct many sentences for English/grammar structure.  The content overall is appropriate for AMT.  My comments above and specific comments below probably warrant a recommendation of major revisions, after which the article -- whose topic is important and intersesting -- will be more useful to the community.

Specific Comments:

First line in abstract: It is unusual to say LH is the essential factor driving convective systems.  It is also a product of convection.  I would reword to say it is an essential factor connected to convective system circulations.

Line 30: convection is definitely not resolved explicitly at a few kilometers.  Over a decade has passed since the 2011 paper cited, and this is appreciated even more.  Perhaps mesoscale convective effects become resolved, but at this resolution, convection is “permitted.” 

L72: recommend rewriting first part of sentence to: “These products provide instantaneous heating estimates, but their temporal resolutions are low compared…”

L75-80: Can this sentence by written differently?  I’m not sure what is trying to be conveyed as written:  “Cloud top information from geostationary data is included when creating cloud analysis during data assimilation (Benjamin et al., 2016), and thus LH retrieved based on cloud top temperature, can be useful in the forecast model by keeping consistency of retrieved LH with the updated cloud analysis.”

L320-322: details about converting units is probably unnecessary information in the article: “but provided in different units. LH from GOES-16 and NEXRAD are in K/s to easily match with modeled heating rate, while DPR products are in K/hour. Therefore, LH in K/hour from DPR products are converted to K/s for comparison.”

L352-353: what is the reason for interpolating to the WRF grid at this point?  WRF will rarely get the convection in the exact right place at the right time as observations due to a different surface relative to reality, so why this is done is not clear to me. 

L369-370: changes in buoyancy are related to the vertical derivative of heating, not absolute heating rate at any level.  If heating is increasing with altitude, then there is a dampening effect on buoyancy specifically.  So, it is not about buoyancy here; convection can be initiated because LH increasing with height induces surface convergence which  favors convective initiation (which the authors happened to mention near the introduction near L34).  Clarify this statement. 

L374-376: some papers using ARM radars (papers led by Die Wang) indicate that 40 dBZ is a good proxy for convection overall.  From that perspective, 28 dBZ is too low too.

L450-459: I do not understand Fig. 7.  For total LH, are you simply summing up all the LH for each convective pixel?  So, if for example there are 100 convective pixels, the total LH is the straight sum across all 100 and not an average?  Or is total LH some combination of convective for every product + stratiform from CSH?  In the literature,  total LH as typically reported as the combination of all convective + stratiform + non-raining + anvil LH.  If Fig. 7 “total” (and Table 4) is calculated simply by summing convective LH, then I do not understand the value of this figure.  It would largely be reflecting differences in convective area and not LH profiles since the previous figures suggested similarity in LH profiles for GOES and CSH.  Instead of summing LH, I would strongly recommend showing the convective area differences for each box and product, and then the reader will be able to infer an overall difference in LH as a combination of both convective area differences as well as profile structure differences. 

L510-511: initiating convection certainly depends on structure and not “total” heating.  Is there a reference to support that the impact on initiating convection is similar if the total heating is similar?  And, again for clarification, what is meant by “total” here – column integrated, or the sum of all individual convective heating profiles? 

L512: the text: “and some conditions that are used…are applied before the comparison to be useful for the real application” is very vague.  I do not know what this sentence means. 

L529-530: there is a lot of scatter in Fig. 8, and I am uncertain again on what is being shown: the vertical integral (“total”) or different levels all combined into a scatterplot of the “total” LH shown for convective regions of Fig. 7?  Perhaps the authors should think about doing some averaging of convective LH over the identified convective regions for each rainfall system and then aggregating those convective-system averages and comparing?  There is almost too much noise to infer anything from Fig. 8, despite the 0.83 correlation. 

L565-600: Fig. 9, NEXRAD (9c) shows convection in the yellow box, but it disappears when GOES LH was used.  Why?  I recommend commenting on this.  Also – is convective LH assimilated everywhere or only in the 4 boxes? 

L670-672.  “Even though one might think… it is actually more than just one brightness value.”  This sentence is too informal for a publication, and should be re-written. 

This article describes a latent heating (LH) estimation method that can be applied to geostationary (GEO) satellite observations. The method consists of two steps: 1) convective cloud detection using temporal changes and spatial patterns in visible reflectance and IR brightness temperature (TB), and 2) once a convective cloud is identified, LH profile assignment based on IR TB by searching a lookup table (LUT) created using WRF model simulations. The authors compared the GEO-based LH retrievals (applied to GOES-16 hourly measurements) with LH from NEXRAD and GPM CSH. Similarities and differences are documented. Finally, they examine the impacts of GEO-based LH on precipitation forecast by applying LH to initiate convection. Forecast results are similar to the one using the NEXRAD LH.

General comments:

Overall, this is an interesting and useful study that has great potential in various applications, because GEO provides much higher spatial and temporal coverages than space-borne and ground-based radars. The proposed method and data product should contribute to convection and precipitation forecasting. For this reason, I’d like to see this paper published.

I only have one major point: the GOE-based LH estimation method described in this study contains two steps: convective cloud detection and LH retrieval for the detected convective cloud. The first step uses a lot more information than the second step. For convective cloud detection, they used multiple channels and their temporal change and spatial structure. In contrast, for LH retrieval, they only used a single piece of information, namely, 11.2-um TB. I wonder if they can consider adding more predictors in their LUT for retrieving LH profiles, given that there are such observations around. Temporal change in TB is an obvious candidate. Meanwhile, environmental parameters will also help. For example, the ambient sounding profile or CAPE has bearings on convective intensity, which should affect the magnitude and vertical structure of LH.

Specific comments:

(Figure 1) The total LH: are they vertical integrals of the LH profile? Can we really integrate LH this way? If LH is 1K/hr at one level and 2K/hr at another level, do we simply add them up?  Some clarification is needed.