Meteorological and Land-Cover Controls on Grassland Fire Behaviour by WRF v4.4-SFIRE v0.1 with the computational optimization

Lou, Mengjie; Wu, Qizhong; Wang, Yongli; Zhang, Baogang; Tao, Jinhua; Gan, Pu; Cheng, Huaqiong; Zhang, Jiating; He, Jiahuan; Fan, Meng

doi:10.5194/egusphere-2026-3195

Preprints

https://doi.org/10.5194/egusphere-2026-3195

Preprints

22 Jul 2026

| 22 Jul 2026

Meteorological and Land-Cover Controls on Grassland Fire Behaviour by WRF v4.4-SFIRE v0.1 with the computational optimization

Mengjie Lou, Qizhong Wu, Yongli Wang, Baogang Zhang, Jinhua Tao, Pu Gan, Huaqiong Cheng, Jiating Zhang, Jiahuan He, and Meng Fan

Abstract. Under global warming, grassland wildfire risk has increased substantially. The WRF-SFIRE model was applied to the extreme grassland wildfire that occurred in Inner Mongolia in 2023, with satellite-derived fire perimeters used to evaluate the effects of meteorological forcing, land cover, and key driving factors on fire behavior, as well as to optimize model performance. For meteorological forcing, the FNL 6 h two-way coupling configuration achieved the highest spatial agreement with observations, with a recall of 39.1 %, whereas ERA5 was more sensitive to short-term meteorological variability but tended to overestimate fire spread. For land cover, FROM_GLC30 (2017) showed relatively high overall agreement but tended to overexpand the fire, GlobeLand30 (2020) produced more conservative simulations, and GLC_FCS30D (2022) achieved the best overall performance. Fire-atmosphere feedback significantly enhanced fire spread rate and burned area. Wind speed exerted the strongest influence on burned area and spread distance, relative humidity mainly controlled spread rate and flame length, and air temperature played a secondary role. Among the fuel-related parameters, fuel load mainly affected flame intensity, fuel depth dominated burned area variation, and fuel moisture effects mainly regulated spread rate. In addition, the optimized computational configuration, particularly the combined use of PNetCDF and asynchronous I/O, reduced runtime by 29.3 % relative to the baseline case. These results clarify the differential effects of input data and key driving factors on grassland fire behavior simulations and provide more reliable support for fire risk warning and management.

Received: 03 Jun 2026 – Discussion started: 22 Jul 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 1573 KB)

Supplement (755 KB)

Download & links

Mengjie Lou, Qizhong Wu, Yongli Wang, Baogang Zhang, Jinhua Tao, Pu Gan, Huaqiong Cheng, Jiating Zhang, Jiahuan He, and Meng Fan

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-3195', Anonymous Referee #1, 23 Sep 2026
General comments
The manuscript does not have a clearly stated scientific hypothesis or a sufficiently focused model evaluation question. Instead, it combines three rather different activities: comparison of meteorological and land cover inputs, one at a time parameter perturbations, and HPC/runtime optimization.

I am not convinced that the experiments support several of the main conclusions. In particular, the model simulations do not reproduce the dominant observed direction and geometry of fire spread but the manuscript repeatedly describes some configurations as having good agreement or improved representation.

The observational basis for evaluation is weak. The “observed perimeter” appears to be a burn scar derived from pre-fire and post-fire satellite imagery, not a time-resolved perimeter. This raises a fundamental question about comparisons made to the purported observed fire extent at 13:00 UTC.

The sensitivity experiments largely demonstrate responses that are already built into the model formulation, and the claimed ranking of “dominant” controls is difficult to interpret because the perturbations of wind speed, RH, and temperature are not normalized or otherwise made comparable.

The computational optimization section seems only loosely connected to the fire behavior science. Some of the I/O and decomposition results may be useful computationally but benchmarking specific Intel compiler versions on one platform does not by itself constitute a substantial advance in modeling science.

Overall, I do not yet see a clear new scientific result. The paper contains potentially useful case-specific observations, but the central evaluation and sensitivity conclusions are substantially stronger than the evidence presented.

Specific comments
Opening motivation / grassland fire risk

The first sentence of the abstract states that “under global warming, grassland wildfire risk has increased substantially.” I do not think this claim is necessary to motivate the study and the cited material does not appear to establish that general statement for global grassland wildfire risk. The paper is about evaluation of WRF SFIRE for one Inner Mongolia event; the motivation should be tied directly to that problem.
First sentence of the abstract

The statement that WRF SFIRE was applied using satellite-derived perimeters “to evaluate the effects of meteorological forcing, land cover, and key driving factors … as well as to optimize model performance” does not follow logically. Satellite observations can be used for model evaluation but they have no direct connection to compiler selection, domain decomposition, or I/O optimization. The abstract combines two substantially different objectives.
Nature of the satellite observations

The manuscript describes an image from 26 April and a post-fire image from 10 May from which the burned area was extracted using BAI and then visually refined. That is a final burn scar not an observed time sequence of fire perimeters. Yet Fig. 5 compares simulations with an “observed burned area perimeter at 13:00 UTC.” What evidence establishes that this final burn scar represents the fire extent at 13:00 UTC? If there is an independent perimeter observation at that time, it needs to be described. Otherwise this comparison is not a valid evaluation of fire progression.
Recall = 39.1%

The abstract gives a Recall value of 39.1% without defining what that is. More importantly, Recall is later defined as overlap area divided by observed burned area. This doesn’t penalize simulated burned area outside the observed scar. Recall alone is not a sufficient metric of spatial agreement. Calling the 39.1% case the “best agreement in both extent and location” seems stronger than this metric supports. Some measure of false positive area / precision or an intersection-over-union type measure would be more informative.
The simulations do not reproduce the observed spread direction

Fig. 5 is, to me, a major problem. The observed burned area has a substantial lateral/easterly extension, whereas the simulations predominantly spread northward/northwestward. Even the best scoring simulation does not reproduce the dominant observed geometry. This should be treated as a model failure to understand, not mainly as a basis for ranking the simulations. In a coupled fire-atmosphere model, getting the fire-driving flow wrong is central to the evaluation. The statement that the FD simulations “more closely reproduce the observed fire geometry and spread pathway” needs to be reconsidered in light of Fig. 5.
What is the actual hypothesis or model evaluation question?

The introduction states three aims: identify suitable input datasets, identify controlling factors, and improve computational efficiency. These are tasks rather than a clearly stated scientific hypothesis or evaluation question. What, specifically, is being tested? For example, is the hypothesis that one meteorological forcing represents the observed fire-driving flow better? That two-way coupling improves simulation of the observed spread? That a particular land cover representation improves the simulation for a physically identifiable reason? The experiment design should follow from a stated question.
FNL versus ERA5

The manuscript correctly states that both FNL and ERA5 are used at 0.25° spatial resolution; their principal stated resolution difference is temporal, 6 h versus 1 h. However, differences between FNL and ERA5 cannot simply be attributed to temporal resolution. These are different analysis/reanalysis systems with different atmospheric states and data assimilation. The FNL–ERA5 comparison a not a clean temporal resolution experiment. The manuscript states that ERA5 produces larger burned areas and higher spread rates and concludes that higher temporal resolution is “more favorable for representing fire spread.” That conclusion does not follow from the comparison. A larger or faster simulated fire is not necessarily a better simulation.
Meteorological sensitivity experiments

Wind speed is perturbed by ±2 and ±4 m s⁻¹, RH by ±5 and ±10 percentage points, and temperature by ±2 and ±4°C throughout the entire domain and forcing period. The manuscript then compares the resulting percentage response of the fire metrics and labels the largest response the “dominant” control. But these perturbations are arbitrary and are not normalized to comparable climatological variability, uncertainty, variance, or nondimensional sensitivity. Thus, conclusions such as “wind speed was the dominant meteorological factor” are conditional on the perturbations chosen. They should not be presented as general physical rankings.
Air temperature sensitivity

I do not understand the interpretation that air temperature “played a secondary role.” What physical pathway is being tested here? SFIRE uses the Rothermel spread formulation, and the manuscript separately enables/disables a fuel moisture model. The authors need to explain how perturbing atmospheric temperature affects fire spread in this configuration and whether that response occurs through fuel moisture, atmospheric dynamics, stability, surface fluxes, or some other pathway. Without that explanation, comparing the magnitude of a temperature perturbation response with wind or RH does not establish the relative physical importance of temperature.
Are some of the sensitivity results simply consequences of the equations?

Several reported findings appear largely to reflect the model formulation itself: increasing wind increases spread; modifying fuel load changes intensity; modifying fuel depth changes behavior according to the Rothermel fuel representation. The manuscript needs to distinguish between a genuinely emergent coupled-model result and a response already prescribed by the fire spread equations. Otherwise, it is difficult to determine what was learned from the experiments.
Land cover comparison is confounded with fuel model mapping

The three land cover products are not merely alternative spatial maps. They are converted to different numbers and distributions of Anderson fuel classes: two produce four mapped fuel classes, whereas GLC_FCS30D produces nine. Consequently, differences between the simulations are not solely “effects of the land cover dataset.” They also reflect the chosen land cover to fuel mapping. This distinction is especially important because the paper draws conclusions about which land cover dataset performs best.
Inconsistent characterization of the best land cover product

The manuscript says FROM_GLC30 gives better overall spatial agreement with the observed burn scar, while GLC_FCS30D better identifies fine-scale nonburnable patches. The abstract nevertheless states that GLC_FCS30D achieved the “best overall performance.” What metric defines “overall performance”? That needs to be stated consistently.
Fire-atmosphere feedback

The manuscript concludes that fire-atmosphere feedback significantly enhances spread and burned area. But in the time series the FD and noFD cases cross, and later the noFD burned area can exceed the FD area. The physical interpretation therefore seems more complicated than simply “feedback enhances spread.” The paper should distinguish time-dependent and process-dependent effects rather than collapsing them into a single statement.
Computational optimization

I understand what is being done: compiler comparisons, MPI versus MPI+OpenMP configurations, grid decomposition, and NetCDF/PNetCDF/asynchronous I/O tests. The question is what scientific or generally transferable result comes from these experiments. The compiler differences are only a few percent and are specific to the Intel versions and hardware tested. The conclusion that compiler versions should be benchmarked on the target architecture is unsurprising.
The PNetCDF/asynchronous-I/O and decomposition results may be more useful. If these are intended as a contribution to GMD, the manuscript should make clearer what is specific to WRF SFIRE, why the behavior occurs, and how transferable the recommended configurations are to other domains, processor counts, and architectures.
Claim of operational relevance

The manuscript repeatedly connects these results to early warning, emergency response, and operational simulation. That claim seems premature when the simulated fire does not reproduce a major feature of the observed spread direction and the observational evaluation itself is uncertain. Computational speed is only useful operationally if the forecast has demonstrated physical and predictive skill.
Technical corrections / presentation
These are mostly presentation issues rather than scientific ones:
Define Recall at first use in the abstract or preferably replace it there with a more readily interpretable description.

Remove most experiment/configuration names from the abstract. Names such as FNL 6 h (FD), FROM_GLC30, and GLC_FCS30D obscure the scientific message.

“with the computational optimization” in the title is awkward English and should be revised.

Use consistent terminology for burn scar, burned area extent, and fire perimeter. They are not interchangeable.

Clarify exactly what is meant by an “observed perimeter at 13:00 UTC.”

The abstract and conclusions should avoid “best overall performance” unless a specific aggregate performance criterion is defined.
Citation: https://doi.org/10.5194/egusphere-2026-3195-RC1
- AC1: 'Reply on RC1', Qizhong Wu, 29 Oct 2026
  
  The point-to-point responses is at the attached supplement.
  
  Citation: https://doi.org/10.5194/egusphere-2026-3195-AC1
RC2: 'Comment on egusphere-2026-3195', Anonymous Referee #2, 03 Nov 2026

The manuscript uses WRF-SFIRE to simulate the 2023 Chenbarhu grassland fire and examines the effects of meteorological forcing, fire–atmosphere feedback, land-cover-derived fuels and several meteorological and fuel parameters. It also presents a set of runtime tests. I think a well-evaluated simulation of this large steppe fire would be useful. At present, though, the evaluation is too limited to support the main conclusions. Fire-spread evaluation relies on the final scar and surveyed final area, no comparison of the simulated meteorology with observations is presented, and several proposed mechanisms lack supporting diagnostics. Some important configuration details are also missing.
I have read RC1 and the authors’ response. My comments focus on what remains unresolved and on the evidence needed to support the explanations given in the response. I think substantial revision is needed.
Major comments
Wang et al. (2023, Fire) already compared coupled and uncoupled simulations of a fire in the grassland–forest transition zone near the Mongolia–Inner Mongolia border, using the same atmospheric grid spacings and FROM-GLC-derived fuels. What does the present study add? I feel this needs a clearer answer in the Introduction. The manuscript already introduces the FireFlux and FireFlux II evaluations, but why are the present coupling results never compared with those studies? Section 3 contains no literature citations. Rothermel (1972) and Coen et al. (2013) are also absent, despite their relevance to the model description.
I could not find complete namelist.input in the materials provided. L163–165 state that the fire-driving wind is interpolated to 6.096 m. Does this match the actual fire_wind_height setting, and why was this height chosen for the grassland fuels? Was an additional wind-reduction factor applied? Please also identify the fire module, spread formulation and effective upwinding scheme. Given the sharp fuel transitions and non-burnable patches, were numerical propagation errors near fuel boundaries checked?
What is the source of the ignition time and coordinates at L124–126? Were 07:30 UTC and 119.09°E, 49.20°N used directly for model initialization? I would also like to know the ignition geometry, radius and duration.
How exactly was wind speed perturbed in Table 4? Which vertical levels, nested domains and input files were changed? Please show how much the wind driving the fire actually changed. Can the wind–slope explanation at L405–407 also be quantified?
What fuel moisture content was prescribed? Was the default 0.08 used, or were different values assigned? AC1 explains that RH affects spread indirectly through changes in the simulated atmosphere while fuel moisture remains fixed. I can follow that argument, but I do not see the diagnostics needed to explain the 85.4% range in Fig. 8c. How much did the fire-grid winds change, and where? For the temperature perturbations, was specific humidity or RH held fixed?
Why is no meteorological evaluation presented? The paper quotes weather observations, but I could not find a comparison of the simulated wind, temperature or humidity with the Hailar station or other available measurements.
How are spread distance and flame length calculated, over which part of the fire, and at what time? The ROS curves in Figs. 4b and 7b look like spatial averages. Are they perimeter averages? “Variability” in Figs. 8c and 9b also needs a definition.
AC1 states that 1299 km² comes from a field survey. What is the area of the digitized scar polygon, and which area enters the calculation of recall, precision and IoU? The manuscript mentions BAI and visual editing, but I would like more detail on how the two satellite images were used and how uncertain the resulting boundary is. AC1 also suggests that suppression explains the missing eastern part of the scar. What records or other evidence support that explanation?
Were VIIRS or Himawari-9 active-fire detections checked for this event? If usable detections exist, could they help constrain the fire progression or estimated arrival times? I also think the abstract should acknowledge that the simulations in Fig. 4 end with burned areas several times larger than the surveyed area, together with the limitation that suppression is omitted.
I am not yet convinced by the explanation at L548–550 that feedback enhances spread by strengthening local winds and promoting low-pressure development. Where are those changes demonstrated? AC1 reiterates the explanation that feedback later limits expansion, allowing the noFD area to exceed the FD area, but this also needs evidence. FD minus noFD fields of wind at the fire wind height, pressure perturbation and heat flux around 12:00 UTC and after 17:00 UTC would help. Do the fire heads encounter the northern river or approach the d03 boundary around the crossover?
I have several questions about the runtime tests. Table S2 changes the floating-point model and target flags alongside compiler version, so I do not think the differences can be attributed to the compiler version alone. How variable were the three repeated timings? What domain decomposition was actually used in each hybrid and asynchronous I/O case?
In Table 9, the (15\times8) configuration is about 17% slower than (8\times15), although both have the same η. The paper recognizes a directional dependence, but why attribute it to the north–south orientation of the fire? Similar benefits from asymmetric decomposition and I/O changes have been reported for atmospheric WRF (Huang et al., 2023, Applied Sciences). A comparison with the fire module switched off would help test this explanation. Please also give the history interval and output volume, and confirm that the optimized configurations preserve the simulated fire and atmospheric results.
Which fire-induced circulations is the 300 m grid expected to resolve, and why is Kain–Fritsch active on the 1.5 km domain?
Minor comments
At L65–67, why describe fire-spread models as explicitly simulating chemical processes? Please qualify this statement for the semi-empirical spread formulation used here. The Introduction mentions chemistry in several places (e.g. L65–67 and L102–105), yet the study does not appear to have the chemistry module enabled. Why is that?
I do not think the distinction between WRF-Fire and WRF-SFIRE at L163–165 is accurate as a general statement. Please check this passage.
Why is the burned-area response described as “consistent” at L411–413 when weaker winds and lower temperatures also increase burned area in Fig. 8b?
At L430–431, the fuel-moisture experiment changes ROS by −20.0% and flame length by −4.5%. Why is the response described as mainly affecting intensity-related metrics?
The recommendation of a 90–95% per-node load goes beyond the reported tests, since no 95% case is presented. I also could not find memory measurements supporting the claim of reduced memory pressure at L567–569.
I feel all the figures need another pass for readability. Several maps lack coordinates or scale bars, and some colorbars and line styles are not explained. Please explicitly mark the model ignition point or line on a georeferenced map. The time axes in Fig. 7 should identify UTC, and the “Area” values in Fig. 5c should be labelled as overlap areas.
The text, tables and references need a thorough proofread. For example, “utilizationefficiency” at L482, the Fig. 2 caption is repeated in the text at L153.

Citation: https://doi.org/10.5194/egusphere-2026-3195-RC2

Mengjie Lou, Qizhong Wu, Yongli Wang, Baogang Zhang, Jinhua Tao, Pu Gan, Huaqiong Cheng, Jiating Zhang, Jiahuan He, and Meng Fan

Supplement

https://doi.org/10.5194/egusphere-2026-3195-supplement

Mengjie Lou, Qizhong Wu, Yongli Wang, Baogang Zhang, Jinhua Tao, Pu Gan, Huaqiong Cheng, Jiating Zhang, Jiahuan He, and Meng Fan

Viewed

Total article views: 601 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
383	159	59	601	84	54	56

HTML: 383
PDF: 159
XML: 59
Total: 601
Supplement: 84
BibTeX: 54
EndNote: 56

Views and downloads (calculated since 22 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	164	22	13	199
Aug 2026	19	15	3	37
Sep 2026	45	36	14	95
Oct 2026	128	76	26	230
Nov 2026	27	10	3	40

Cumulative views and downloads (calculated since 22 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	164	22	13	199
Aug 2026	19	15	3	37
Sep 2026	45	36	14	95
Oct 2026	128	76	26	230
Nov 2026	27	10	3	40

Viewed (geographical distribution)

Total article views: 573 (including HTML, PDF, and XML) Thereof 573 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 05 Nov 2026

Short summary

Climate change worsens wildfires, but fire drivers are unclear and models run slowly. Simulating a 2023 Inner Mongolia grassland fire, we tested how weather and land cover affect simulation accuracy and found that wind controls burned area and spread distance, humidity drives spread rate and flame length, while fuel depth dominates burned area. In addition, computational optimizations cut runtime by 30%. These findings deepen understanding of fire mechanisms and improve model emergency response.


Total:	0
HTML:	0
PDF:	0
XML:	0