Dear Dr. Wang, The reviewers still have major concerns. Please carefully consider the comments/suggestions and revise the manuscript accordingly. Best regards, Yafang Cheng

#
The revised version has been improved largely. The remaining comments are about the parameters of XGboost method and the reliability of the RF importance index.

1. Could you describe the parameters of the XGboost method, such as the number of trees, the maximum tree depth, et. al.
2. Although you show the RF importance is stable under cross-validation, only the model with high accuracy presents true physical relationships. Because of some limitation factors, such as the number of samples and the strong non-linear problem, the machine learning method may be stable under a low accuracy, but it can not prove that the importance index is reliable. I think the authors can verify it by physical knowledge. It is also because the machine learning method is black box, and we want to get knowledge by combining physics and machine learning.

#
This manuscript integrated multiple machine learning algorithms and developed a prediction model at the resolution of 0.5ox0.5o to predict monthly wildfire burned area over the South Central United States during 2002-2015. The relative importance of the environmental drivers is also identified. The paper was well structured and presented. It can be accepted for publication if the following questions can be addressed.
1. In Figure 6, both rhum_anomally and rhum are selected as important predictors for the winter-spring fire season. What is the difference between them and are they highly correlated? Also, please show the correlation coefficients between the variables in Figure 6.
2. In section 4, the authors compare the performance of the prediction models using MLR to predict burned area in Europe, Western US, etc. Please add several figures to show the performance of MLR model in the prediction of burned area in South central US using the same variables in Table 1 as a comparison of the model using 4-step machine learning.

#
This manuscript proposes a new method for modeling wildfire burned area that includes a combination of random forest, quantile regression and logistic regression which should help with the issue of unevenly distributed burned area data. This technique was applied for modeling the amount of burned area in the South Central United States during two fire seasons. The author concluded that antecedent climate conditions, in particular relative humidity, is the main driver of the amount of burned area in the winter-spring season while concurrent weather conditions drive fire activity during summer months. However, this proposed method has several limitations that some might argue are substantial and require more attention than being listed at the end of the manuscript. Additionally, the lack of information regarding how this technique can be applied to different regions with longer fire return interval and/or different datasets makes the proposed method more questionable. Moreover, purposely excluding all non-natural fires and out of more than 30 variables selecting only one not related to climate (population density), makes the main conclusion that climate is the main driver of fire activity biased especially considering that most of the fires in the US have an anthropogenic origin. All things considered, I would recommend a major revision.
P1L20 and L22: I would recommend replacing “the magnitude of total burned area” to “the amount of total burned area”
P1L27: You cannot say “many regions of the world” have experienced an increase in fire activity while only citing papers from the US. You should either add citations from different regions of the world or only write about the US.
P2L37-38: Please reword this sentence. “…complex interplay …can change by spatial scale…” does not sound correct.
P3L73: Can you explain the choice of the study area? In the title and all over the paper you claim to model fire activity over South Central US while your actual study area is not a geographic region or a set of ecoregions. It is an arbitrary rectangular. How is that a “vegetation rich” part of South Central US? Is it based on a particular vegetation map? If so, what was the threshold?
Additionally, it will be beneficial to include a paragraph or to describe the study area after the introduction. To explain in more detail why it is important to model fire activity in that particular part of the US since most studies concentrate on California and Pacific Northwest, regions more fire-prone comparing to South Central US. Also a detailed description of its fire regime, vegetation and climate is necessary.
P3L94: Here you provide the accuracy of FPA-FOD for the period 1992-1997. How is this relevant to your study when your study period is 2002-2015? Your study period is covered by the MODIS fire product. Why not to use global fire product instead? Usually, the benefit of using regional data is its accuracy; however, here small fires which I assume are the majority in South Central US are not included. Additionally, most places in the world do not have datasets similar to FPA-FOD. If the results of your proposed method are sensitive to the choice of the dataset does it mean that this method can be only used within the US?
Table1: It is not clear that the source of most of the weather/climate variables is NARR, neither their resolution.
P4L116: You need to cite every dataset that was used in this paper.
P4L122: Citation is needed.
P5L127: “Situations” is not a good choice of words.
P5L155: Citation is needed.
P6L171: What exactly is “ecoregion type”?
P6L174: What is “EPA”?
P7L215: I would argue that 0.7 is a very strong correlation. Lower correlation to 0.5 will help to reduce the number of predictor variables which is way too many and most of them represent similar processes.
P8L252: “Larger” should be replaced by “greater”.
P10L306-330: While I understand that comparison to other work is important, I am not sure that selected studies are comparable. I would argue that there are a lot of studies that modeled the amount of burned area with R2 higher than 0.4. I agree with the authors that funding a study that targeted South Central US is challenging since it is not a region particularly interesting in terms of fire activity. I would suggest applying this method to other regions in the US to provide evidence that this method can be used outside of the study area which has very little fire activity.
TableS3: Why the results are presented for 2011 both seasons together, 2014 winter-spring, 2008 summer season? That is confusing. Can you include the results for each year, each season.
P11L338: The results show that the proposed model performs worse if the amount of burned area is outside of the norm in terms of MAE and RMSE. I would say significantly worse. Doesn’t it contradict your conclusion that the proposed technique can be used to model future fire activity due to the changing climate? If the model cannot predict unusually high fire activity which as the authors mentioned in the introduction is recently observed in the US than how can it be used for future predictions?
P12L369: Can you elaborate on why the predictors explained much less of the variability of the burned area during winter-spring season compare to summer fire season?
P13L391: Is it possible to provide information about the relationships between the amount of burned area and the most important predictors for each fire season. For example, if the relationship between RH and the burned area is negative for both the winter-spring and summer season? I would assume that moisture during the antecedent conditions should be positively correlated with the burned area while an excess of moisture during the fire season will suppress fire activity. It would be interesting to see if those relationships vary depending on the season.
P14L418, L427: In “fire burned area” the word “fire” should be removed.
P14L431: You claim that your large sample size captures the variability in wildfire activity while in L515 under limitations you admit that the model cannot predict burned area greater than it was observed before. These two sentences contradict each other. In reality, it only captures variability within 14 years which is not a long time period in terms of the climate-fire relationship. The assumption that the amount of burned area can never exceed the one that was observed in the past 14 years is flawed and cannot be held according to numerous future predictions few of which the authors cited in the introduction. While the study area is represented mostly by grassland which might have short enough fire return interval to be captured during 14 years, forested regions experience fire every 100 years and longer; therefore, this particular method cannot be used in regions with fire reoccurrence longer than a study period which needs to be clearly stated. And even in grassland, fire activity can change drastically due to climate change, population growth, social-economical changes. This limitation is too important to overcome by simply mentioning it at the end of the paper. While this model can be used to evaluate what environmental factors drive present fire activity which can benefit fire modeling, it cannot be used for future predictions.
P14L445: Please, rewrite this sentence. The burned area is not “contributed by” any controls.
P14L446: I do not understand how all factors can increase the burned area? I would expect that some predicted variables are negatively correlated.
P15L460: Which countries exactly when you only cite studies from the US?
P15L470-473: Add citations.
P15L477: There is no doubt that climate variables will appear the main driver of fire activity considering that fires outside of natural fire seasons were excluded, together with prescribed fires to concentrate solely on environmental factors that do not include any variables other than climate. Other environmental variables that could be included are soil, elevation, slope. But climate anomalies and climate means are both climate variables even if the authors artificially separated them into different categories.
P16L486: I would suggest either find a better way to deal with categorical variables or not including them since they are not treated equally with continuous variables.
In general, section 5.2 is not providing any important information and I would suggest removing it from the paper. All variables are arbitrarily assigned to 4 groups while in reality all of those variables except population density are climate variables to some degree. I would suggest concentrating on specific variables which number can be significantly reduced after accounting for multicollinearity. The effect of each category on the amount of burned area is directly related to the choice of variables and is very subjective.
P16L495: This is not Discussion and Conclusion section. It is Limitations and Conclusion. All the discussion is included together with results.
P16L502: I do not agree that R2 0.4 is high enough to overperform most past fire studies.
P16L507: Here you reported the percentage of the grids with a correlation higher not “larger” than 0.5. This implies that you consider correlation 0.5 and higher significant. In this case, you should use the same threshold for multicollinearity. Or report the percentage of the grids with a correlation higher than 0.7.
P16 L513: As I mentioned before this is a very big limitation and it needs to be discussed and suggestions for its overcoming need to be proposed. Additionally, it is necessary to state how large should be the study area to have enough data to obtain a distribution which will be representative together with how long time series is needed depending on fire regime characteristics of the study area. Failing to convince a reader that all limitations can be overcome or at least in which case the assumptions of the model are valid is the main problem with this paper. It is not clear how this technique can be transformed to a different region with higher year-to-year variability in fire activity comparing to South Central US.

Part of this manuscript is written in the past tense and part in the present tense. Al least within a section you should select one and be consistent.

Dear Authors,

While the manuscript has been improved, there are still a few major concerns need to be addressed. Please revise the manuscript according to the two additional points raised by referee #3 (attached). Also, it seems that one the major concerns about the necessity to use 58 variables (since many of which are highly correlated) from previous round has not been fully addressed. The authors may consider to use some statistical ways to test for multicollinearity, e.g., computing the variance inflation factor (or VIF), rather than choosing 0.7 as a threshold for correlation. In addition, the South Central US has been chosen as a study area where the risk of wildfires has been predicted to be the highest in 2031-2050. But the proposed model seems to fail to predict BA during the years with abnormal fire activity especially during spring fire season (figure 5). If it is because “random forest or quantile regression forest cannot predict burned area greater than it observes before”, how would that influence the performance of the proposed model for future predictions assuming that fire activity will increase in the next several decades. Please comment on it and discuss it in the revised manuscript.

Referee #3

1. Per previous suggestion, the authors give the parameter information of the XGBoost. The different parameter configurations in the XGBoost and Random Forests are used for winter-spring and summer. Perhaps, we want to a uniform robust machine learning model that can achieve high accuracy both in the different seasons.
2. The authors analyze how RH anomaly and temperature anomaly affect the prediction. But the temperature anomaly is just ranked 10th in the summer season. I think the authors need to analyze how the top at least 3 variables affect the prediction so that we can learn something from the machine learning model, not just the accuracy. On the other hand, the author said “The physical reason behind their importance is that higher temperature coupled with lower relative humidity in the summer can cause drier fuel and this condition is favorable for fires to start, spread, and burn more intensely”. But, the machine learning importance cannot provide the influence of change of variable values. The authors should further prove that.

Look forward to the revised manuscript.
Best regards,
Yafang Cheng