Forecasting Live Fuel Moisture of Adenostema fasciculatum and Its Relationship to Regional Wildfire Dynamics across Southern California Shrublands
Round 1
Reviewer 1 Report
The authors did a commendable job of addressing the points in my previous review. I especially appreciate that they took the time to improve the cumulative distribution modeling by converting LFM to rank percentiles.
I do think that one previous criticism has not been addressed. The authors identify what they claim are distinct thresholds and domains, but there’s no proof that the domains are distinct and not just part of a continuous relationship between LFM and fire activity. In ecology, a threshold occurs when a small change in a driver (e.g. LFM) results in a big change in the state of ecosystem (e.g. cumulative area burned). Piecewise regression is designed to find thresholds, but it will also place breakpoints where no threshold exists to maximize fit. In other words, piecewise regression still identifies breakpoints even on a continuous curve lacking a state change.
A good example of the issue I am concerned about can be seen in Figure 4. State changes could be present where there are obvious changes in slope, e.g. at approximate 11 and 24 percentile LFM. However, the piecewise regression is fitting the general shape of the curve, because it is limited to just two breakpoints. Slope on either side of the two identified “thresholds” is nearly identical, and it’s hard for me to see that this represents distinct domains rather than just part of a more continuous variation in the relationship between LFM and cumulative burned area.
This may come down to what the authors mean by using the term “threshold”. My sense from the manuscript is that “threshold” actually means breakpoints identified by piecewise regression, but that the authors interpret this to indicate a state change in fire activity that may not actually exist. I think more explanation of how thresholds should be interpreted could be added to the manuscript, and that statements about the distinctiveness of thresholds and domains should be better supported by evidence.
Author Response
Reviewer 1:
The authors did a commendable job of addressing the points in my previous review. I especially appreciate that they took the time to improve the cumulative distribution modeling by converting LFM to rank percentiles.
I do think that one previous criticism has not been addressed. The authors identify what they claim are distinct thresholds and domains, but there’s no proof that the domains are distinct and not just part of a continuous relationship between LFM and fire activity. In ecology, a threshold occurs when a small change in a driver (e.g. LFM) results in a big change in the state of ecosystem (e.g. cumulative area burned). Piecewise regression is designed to find thresholds, but it will also place breakpoints where no threshold exists to maximize fit. In other words, piecewise regression still identifies breakpoints even on a continuous curve lacking a state change.
A good example of the issue I am concerned about can be seen in Figure 4. State changes could be present where there are obvious changes in slope, e.g. at approximate 11 and 24 percentile LFM. However, the piecewise regression is fitting the general shape of the curve, because it is limited to just two breakpoints. Slope on either side of the two identified “thresholds” is nearly identical, and it’s hard for me to see that this represents distinct domains rather than just part of a more continuous variation in the relationship between LFM and cumulative burned area.
This may come down to what the authors mean by using the term “threshold”. My sense from the manuscript is that “threshold” actually means breakpoints identified by piecewise regression, but that the authors interpret this to indicate a state change in fire activity that may not actually exist. I think more explanation of how thresholds should be interpreted could be added to the manuscript, and that statements about the distinctiveness of thresholds and domains should be better supported by evidence.
Response: text was modified throughout the discussion section to better address this point. Specifically, the following was added to address these issues (lines 399-491):
We find that LFM plays an important role in the spatiotemporal distribution of fire at landscape to regional scales, supporting previous evidence that fire is more prevalent at times and locations in which LFM falls below certain thresholds [2,3,5,6]. Independently, the relationship observed here between LFM and the average size of wildfires (Fig. 5) demonstrates that lower LFM is conducive to greater spread by individual fire events. These results confirm the findings of previous studies that found more rapid spread of fire among low-moisture fuels in laboratory experiments [14,38]. Further, studies in the Chaco Serrano subregion of Argentina have also observed that high LFM often constrains the size of fire events [6], resulting in a greater frequency of large fires during periods of time when LFM is low. Similar patterns have also been found throughout Mediterranean shrublands, where low LFM among shrub species has been associated with high burned area and greater frequency of large fires [4]. Thus, low LFM is strongly associated with both greater ignition success and initial establishment of wildfire events (Fig 5), as well as a greater likelihood that each individual fire event will, once established, spread successfully to cover a wide area. It should be noted, however, that the thresholds identified in this study represent optimal breakpoints as identified through segmented regression, and do not necessarily imply sudden or drastic state changes in LFM~wildfire relationships. Instead, they should be interpreted as marking general transition points in the different phases of the fire season and in wildfire dynamics, as ignitability and fire spread rates change with progressively drier fuels. These dynamics will also likely be affected differently by local topography, fuel structure, microclimate, seasonal wind patterns, and many other factors that may impact the relationship of LFM to local wildfire dynamics [39,40]. Nevertheless, thresholds of dry fuel moisture have been implicated as the triggers for dynamic transformations of forest flammability across multiple continents and vegetation types [5,41]. Thus, as live fuel also makes up a significant component of the flammable material across the landscape, it is likely that the changes in LFM will also impact wildfire behavior.
Our finding of two thresholds in the relationship between simulated LFM and both burned area, mean fire size, and cumulative number of fires indicates that there are multiple domains of LFM~fire dynamics for this shrubland system. These finding support previous studies that identified multiple LFM~wildfire domains across both Europe and northern Africa [42]. The first domain represents conditions in which fuels are moist, and as a result, ignition rates are low, fire sizes remain small, and cumulative burned area remains low. As fuels remain sufficiently hydrated as to limit both ignitions and fire spread throughout this domain, changes in LFM within this domain appear to have only minimal effects on ignitability, fire size, or rates at which burned area increases. There is a wide range of variation in the LFM that bounds this first domain (~77%-96%) depending on which metric is being examined, indicating that this transition point is likely somewhat gradual, and may occur at slightly different LFMs depending on the aspect of wildfire being measured. The second domain represents the range of LFM in which fuels are becoming sufficiently dry for progressive reductions in LFM to impact ignitability and fire spread more strongly. Throughout this middle domain, progressive reductions in LFM appear to be associated with moderate increases in both ignitability and fire spread rates (as measured through examinations of mean fire size). However, these increases begin somewhat gradually in the upper portion of this domain, and only produce systematic increases in the rate of cumulative burned area accumulation throughout the lower half of this domain (when conditions range from ~62%-~77% LFM, Figs. 4, A2). The third and most distinct domain represents the driest conditions, when any reduction in LFM results in dramatic increases both to ignitability (as measured by cumulative number of fires) and fire size. This domain consists of conditions when LFM falls below ~62%, and represents conditions in which fuel is extremely dry. In this domain, both fire frequency and size increase dramatically with reductions in LFM relative to conditions experiencing LFM above this threshold, leading to rapid increases in cumulative burned area, mean fire size, and fire frequency. In contrast to the transition between the first two (moister) LFM domains, which experienced a comparatively gradual transition in the relationships between LFM and ignitability and fire spread, this domain exhibits sudden, sharp increases in both cumulative number of fires, fire size, and cumulative burned area once LFM falls below ~62%, particularly when viewed as actual LFM rather than LFM percentile (Fig. A2). Thus, this lowest LFM threshold likely represents a critical indication of hazard for high fire conditions, as both fire frequency, mean fire size, and cumulative area burned were observed to increase sharply below this threshold.
Previous examinations of LFM in Los Angeles county identified only a single sharp threshold in the relationship between LFM of chamise and cumulative burned area [3]. This observed threshold coincides with the observed threshold between the wettest and middle domains in the LFM~burned area relationship (at 77% LFM) that was observed in this study, although examinations of the relationships between LFM and mean fire size or cumulative number of fires exhibited some variation in the placement of this threshold (77% - 96% LFM). The sharpest and most consistent threshold observed in this study, however, occurred under much drier conditions, at ~61%-62% LFM, representing the transition point into the third and driest domain observed in this study. This lower threshold also corresponds to perceived LFM thresholds often used by firefighters to determine local fire danger throughout southern California [15]. Previous examinations of fire size distributions have also found that fire sizes across many California ecosystems exhibit a three-domain distribution, which supports the results of this study [43]. These findings also indicate that LFM may be the driver (or one of the primary drivers) in generating such a distribution of fire sizes. Inconsistencies in the upper LFM threshold (~77% - 95%) may be partially due to the greater error in predictions of LFM during conditions when fuel moisture is increasingly wet, and may also represent a more gradual transition between the wettest and middle LFM domains. Additionally, some aspects of differences in LFM~wildfire relations among these domains may be influenced by seasonal differences in wind and the confluence between the timing of progressive reductions in LFM and the occurrence of Santa Ana winds and other extreme events that are closely associated with extreme fire sizes and rapid fire spreads [44]. As LFM predictions in this study were driven by local climate conditions, many effects of long-term climate shifts (such as increasing temperatures or drought conditions ) that may have occurred across the study period should be reflected in predicted LFM values, it should also be noted that this study does not explicitly account for any potential nonstationarities across the study period that may have resulted from changes in land use or anthropogenic ignition rates. Thus, despite this strong concurrence with commonly used LFM thresholds based on expert information [15], it is not yet known how generalizable our findings are to other regions or vegetation types.
Reviewer 2 Report
The authors have attended to all my suggestions and, after reading all the answers presented to the other Reviewers, I consider this paper of good quality and ready to be published
Author Response
No comments requiring response in this revision
Reviewer 3 Report
This manuscript aimed to predict the LFMC for shrubland in California and identify the critical threshold of LFMC that is likely to trigger fire when below these thresholds. The topic is interesting and it is better than its previous version. I recommend the publication and have some minor comments:
1. Checking references 7 and 17. They are the same.
2. The authors applied the GAM method to estimate the LFMC, but the accuracy is not that promising despite a reasonable R2=0.654. But the LFMC is slightly overestimated. The authors may have a try on the recently published Global LFMC dataset via (https://firewatching.cn/world_FMC/) in the future.
Xingwen, Quan, Marta Yebra*, David Riaño, Binbin He*, Gengke Lai, Xiangzhuo Liu. Global fuel moisture content mapping from MODIS[J], International Journal of Applied Earth Observation and Geoinformation,2021,101:102354.
Brown, T.P.; Inbar, A.; Duff, T.J.; Burton, J.; Noske, P.J.; Lane, P.N.J.; Sheridan, G.J. Forest Structure Drives Fuel Moisture Response across Alternative Forest States. Fire 2021, 4, 48.
Dennison, P.E., & Moritz, M.A. (2009). Critical live fuel moisture in chaparral ecosystems: a threshold for fire activity and its relationship to antecedent precipitation. International Journal of Wildland Fire, 18, 1021-1027
3. Figure 1 is not readable. Legend is needed and this figure also needs to be polished.
4. Figure 2 should have the same X and Y scale.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Round 2
Reviewer 1 Report
The authors addressed my remaining concern, and the manuscript should be accepted.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
This manuscript is an interesting examination of relationships between live fuel moisture and fire activity in Southern California. There are a few major and several additional issues that should be addressed.
Major issues:
1. The manuscript is not up to date on the current “state of the science” of LFM estimation and forecasting. The authors argue that there is little work done in this area, but need to examine recent work such as:
Capps, S. B., Zhuang, W., Liu, R., Rolinski, T., & Qu, X. (2021). Modelling chamise fuel moisture content across California: a machine learning approach. International Journal of Wildland Fire.
Zhu, L., Webb, G. I., Yebra, M., Scortechini, G., Miller, L., & Petitjean, F. (2021). Live fuel moisture content estimation from MODIS: A deep learning approach. ISPRS Journal of Photogrammetry and Remote Sensing, 179, 81-91.
Vinodkumar, V., Dharssi, I., Yebra, M., & Fox-Hughes, P. (2021). Continental-scale prediction of live fuel moisture content using soil moisture information. Agricultural and Forest Meteorology, 307, 108503.
2. I have concerns that the modeled LFM values produce very different thresholds than the observed LFM values in the analysis shown in Figure 4. Why is this happening? Is error in the GAM model potentially responsible for obscuring actual thresholds?
3. Quantification of error and uncertainty needs to be handled better. GAM modeling error quantification is limited to r-squared, but it’s unclear what kind of error in estimated live fuel moisture this translates to. Presumably the error from the GAM model would then translate to uncertainty in predictions of fire activity? This could be better addressed.
4. In Figures 4b, 5a, and 5b, it’s not clear that there are actually thresholds present in the data shown. The curves appear continuous. Why is a piecewise linear regression appropriate for these more continuous relationships? Is using this type of model producing thresholds that have no or minimal meaning? A good example is the 83% threshold in Figure 4b. I would have a lot of difficulty distinguishing differences in the relationship between LFM and cumulative area burned on either side of this threshold, and yet different domains are defined based on it.
Also, Pimont et al. (2019) warn that breakpoints can appear in cumulative data with no apparent relationship between fire size and LFM, based purely on the statistical distribution of fires across the range of LFM values. You cite Pimont et al. (L81-85), but don’t resolve the questions their findings raise for your dataset.
5. The manuscript uses live fuel moisture and climate variables at a monthly resolution. This should be acknowledged as a limitation of modeling, along with assessment of the variation of live fuel moisture within single months. If a fire occurs at the beginning or end of a month, how different could live fuel moisture be during the fire compared to the monthly mean?
6. The term “forecast” is used frequently, but how the modeling presented in the manuscript could be used in a forecasting framework could be better explained. This is particularly true in light of Capps et al. How would a monthly forecast of fire size and frequency be used? What kind of uncertainty would be associated with forecasts of these parameters? Is “near real-time” estimation of LFM from climate data (L341) useful at monthly resolution?
7. The statistical analyses in the manuscript could be better explained. I understood modeling cumulative burned area, but not modeling of fire frequency (see the comment on stationarity below) or fire size/large fire proportion.
Additional comments:
Inconsistent subsection numbering. Sections are numbered, but most subsections are not. A set of figures gets its own numbered subsection (3.2). Also, it’s unclear why figures get their own subsection rather than being in the subsections that cite them.
L15-16: “However, spatially explicit forecasts of LFM are rare, and LFM’s relationship to regional wildfire dynamics remains largely unknown.” This doesn’t seem to be accurate. Capps et al., 2021, Zhu et al. 2021, and Peterson et al. 2008 are a few examples of spatially explicit models. Similarly, LFM’s relationship with wildfire has been previously characterized.
L20: Should this be “chamise LFM is strongly associated”?
L22: verb tense disagreement: “increase” and “fell”
L40-41: Left margin needs to be fixed
L59-65: This section needs to be updated to include the recent work as described above. Does the fact that the Capps model appears to be used operationally go against the point that current models are “quite limited”?
L88-89: Can you cite examples?
L133: Doesn’t indicate the type of rescaling done. Presumably some type of spatial averaging?
L160: Capitalization
L161-169: Can you address issues with the FVEG and EVT maps representing one point in time vs. actual vegetation dynamics? This would seem to be especially important with fires spanning 66 years.
L178-179: Confusing sentence.
L183: Based on “observed” at the start of this paragraph, I assume that 1952 can’t be right. Don’t the LFM measurements only go back to 1977?
Figure 1: What does the background map show? Why show all of California, when the area of interest is Southern California? Can individual fires be made more clear? There’s no way to see 1818 fires in this figure, since most of vegetated Southern California appears to have burned in the 1952-2017 period.
L185, 191, 197: Do these need to be separate sections?
L192-195: Is there a stationarity assumption for fire frequency? I don’t understand how cumulative number of fires translates to fire frequency, considering that fire frequency likely varies greatly over the length of the record.
L185 & 197: Based on the subsection titles, it isn’t clear how “burned area” and “fire size” represent different quantities.
L201-202: Mean size of what population of fires?
L202: Previous question about fire frequency also applies here.
Figure 3: Define what solid red and dashed-blue lines represent.
L244: Are these actually strong thresholds, or just the best breakpoints when a piecewise linear regression is applied to a continuous data distribution?
L329: “sufficient accuracy” This accuracy needs to be quantified.
L331-333: Can you provide quantitative estimates of error to support this statement on accuracy?
L351: I understand the temporal distribution, but it seems like the spatial distribution is underdeveloped. Can you describe how your analysis provides insight on spatial distribution of fires in the study area? Or is this statement purely based on fire size?
L366-409: The identification of these “distinct domains” is based on the application of a piecewise linear regression to what appears to be a continuous data distribution. I’m not confident how distinct these domains actually are.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 2 Report
This paper, “Forecasting live fuel moisture of Adenostema fasciculatum and its relationship to regional wildfire dynamics across southern California shrublands,” aims to evaluate the role of live fuel moisture in chamise-dominated ecosystems in southern California, how changes in live fuel moisture can be predicted from climate data and how they are related to fire frequency and size. The paper provides important insights into fire ecology for chamise-dominated ecosystems in southern California.
While the authors do a good job of explaining the details of the study and the analysis, I do have a few major concerns. First, the manuscript structure would benefit from some revisions focused on clearly linking the pieces together for the reader. Specifically, authors should consider making sure research objectives (or questions) are outlined in the final paragraph of the introduction (e.g., We have X-number of objectives and then spell them out). Ensure the objectives are linked through the remaining sections to the type of analyses outlined in the methods, the results sections, and the discussion. Second, the methods could benefit from a study area description to set the stage for readers that may not be familiar with the area. Third, critical pieces of information seem to be introduced in the results but should be clearly outlined in the methods as part of the analysis. For example, in line 263, specific areas are mentioned, but I could not find the reason for this in the methods. Forth, it seems a bit odd to me that there is a separate section for figures and tables. These should be placed appropriately throughout the paper, not all designated to a section. Please look at other publications in FIRE for examples. Lastly, the discussion would benefit from subsection headers that clearly link the research objectives and the results. I think if the authors are able to implement these changes, that manuscript would be greatly improved and make an important contribution to the field of fire ecology.
Line Items
39-41 Split into two sentences.
75 what does “critical LFM threshold for fire activity of ~79%” mean? Do you mean when LFM is at 79%? Please clarify.
89-91 Do you mean that there is a temporal disconnect? This sentence is challenging to follow.
92-95 This sentence is challenging to follow. What is the observational limitation? Maybe split into two or reorganize.
76-78 This sentence is confusing. I don’t understand how more area can burn than is available.
140-141 What are the mean conditions over the preceding six months?
156-158 I think this sentence needs a bit more explanation. How did you choose 0.02 for the exclusion? What do you mean by unnecessary complexity?
208-209 Can you reframe this sentence to remind the reader what the research objective was?
213 Acrynyms were used above, but now you are lumping variables together and using words; please clarify. Also, replace “to exhibit” with “exhibited.”
224 The word “likely” is confusing to me in this sentence. Is it likely or does it reflect the dry season? TMAX is associated with a month and certain months are associated with the dry season. Please clarify.
248 Delete duplicate “in”
263 What is the meaning of the two spatial thresholds? I must have missed this in the methods? Please make sure this is clearly defined, i.e., the what and why of the two spatial extents, in the methods
333-335 This sentence is challenging to follow. What portions of the year have high fire risk?
Figure 1. Please remind the reader what years the fire scars span. Can you add a zoomed-in map of the study area with fire perimeters and the distribution of Chamise? The second sentence is a little confusing as written; please revise.
349 What do you mean by potentially no-analog?
Figure 2. Please time range so the figure can stand alone.
Figure 3. Is this figure from your GAM models? Are there units associated with the values? Please add more detail to help the reader understand what you know.
Figure 4. I am confused about how the percent LFM can be more than 100; please explain. Also, please specify which panel the sentence belongs to. For example, the second and third sentences describe panel c. Please spell it out for the reader, so they don’t have to put the pieces together.
Figure 5. Similar comments to figure 4. Please revise. Also, fix the panel labels a and b to align them at the top.
Table 1. Regarding the last sentence, can you remind the reader how data was split for model development versus cross-validation?
Author Response
Please see the attachment
Reviewer 3 Report
In this manuscript, Authors investigate the relationship between LFM and the dynamics of forest fires in seasonally dry environments, namely the size and frequency of forest fires.
The investigation is carried out in a scenario of Adenostema fasciculatum chaparral bushland in an area where wildfires are often intense and spread quickly.
This work is useful and timely opportune since both the published scientific knowledge relating LFM flammability with forest fire dynamics and the existing scientific knowledge about the thresholds in the relationship between burned area and LFM are scarce.
The presented LFM forecasting methodology (which considers a history of climatic data and, based on this forecast, allows assessing the historical relationships of LFM with the size and frequency of forest fires) is adequate. In fact, spatially explicit forecasts of LFM among most vegetation types remain quite limited because the dynamics is complex to model. Remote-sensing methods have been used successfully, but they are often unable to distinguish between LFM of co-occurring species and are incapable of forecasting future LFM. In this context, the methodology presented in this paper is a suitable option once it allows, with very acceptable accuracy, to identify critical phases of fire ignition and spread.
The findings are in line with the scientific knowledge published to date (although the nature and location of thresholds may vary among vegetation types and among regions) and in line with laboratory experiments that corroborate a strong link between fire behavior to LFM among live chaparral fuel beds.
This paper is written very competently, and the presented results are valuable (Based on the results, the authors defend that LFM is strongly associated with the extent, size and frequency of fires (which is not an extraordinary conclusion) and defend that, in the studied scenario, the increase in both the area and the number of fires is drastic when the LFM drops below 63%, which is new and interesting information once it will allow deciding about the implementation of management measures for fire danger control in California).
This work is also valuable once it can be considered as an encouragement for further work in other environments.
My only suggestion for corrections is to proper label Figures (see fig in lines 231, 245, 364 and Figs in line 354)
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 4 Report
see the attachment
Comments for author File: 
Comments.pdf
Author Response
In response to these criticisms, we have added additional information as to the accuracy of LFM predictions across this study (see above) and have corrected our threshold regression analyses to compensate for different frequencies of LFM across the observed area and time periods as recommended by Pimont et al.
