By Mélanie C. Rochoux1*, Cong Zhang2, Michael Gollner2 and Arnaud Trouvé2
1 CECI (Climate, Environment, Coupling and Uncertainties),
CERFACS-CNRS, Toulouse, France
2 Dept. of Fire Protection Engineering, University of Maryland,
College Park, MD, USA
* Corresponding author
Providing accurate predictions of the spread of wildland fires has long been a goal of the fire research community. Whether used as a planning tool prior to prescribed burning or as an operational tool to predict the growth of uncontrolled wildfires, the accuracy of wildland fire spread models and their ability to provide useful information in a timely manner are of paramount importance. This is particularly true in the perspective of changing wildland management practices, the movement of populations to rural areas and climate change, where these coupled influences dramatically increase the risk of highly destructive fires known as “megafires”, with strong implications for public safety and air pollution far away from wildfires.
Despite the development of a number of models, the use of wildland fire spread modeling has been relatively limited operationally. Some of this stems from the fact that all models are by nature approximate, simplified versions of reality (the problem of wildfires is particularly complex to model due to the wide range of relevant spatial scales and to the multiple physical processes involved, ranging from biomass pyrolysis, combustion and flow dynamics to atmospheric dynamics and chemistry). Available data to initialize and parametrize these models, such as fuels, topography, and weather, are also subject to large uncertainties and limited resolution, both spatially and temporally. This was emphasized in the March/April 2016 issue of Wildfire Magazine, where author Rachael Quill highlighted one of the most challenging issues in fire modeling today: uncertainty.
A new approach to this problem is to couple existing models and real-time observations, with the objective of reducing the uncertainties in both model fidelity and input data by using real-time observations of the wildland fire dynamics. This approach is called “data-driven modeling” (or “data assimilation”). Data-driven modeling allows an optimal use of available information and leads to improved forecasts of the system evolution, which we believe holds great potential for the wildfire community. Data-driven modeling thereby offers to take full advantage of the recent advances in remote sensing technology for real-time wildfire detection and tracking. This is critical in the context of climate change, where accurate predictions of the resulting change in the fire regime and intensity cannot rely on past-observed wildfire events. Following a 2015 workshop, “Towards Data-Driven Operational Wildfire Spread Modeling,” we will present here some of the challenges and opportunities this new approach offers.
While real-time data-driven modeling is at an early stage of development for wildfire applications, it is envisioned that this approach will eventually be similar to current weather forecast capabilities, providing real-time fire forecasts including a description of both wildfire dynamics and plume emissions. Currently, weather data such as wind speed, temperature, and cloud cover, from real-time sensors are compared with current model runs and used to improve predictions, compensating for uncertainties and errors in initial data. For fire modeling, the location of the fireline can be used as real-time observation data, where it is then used to adjust simulated fire location, fuels, and even changing weather conditions to improve future predictions of the fireline propagation. While there are challenges to this approach (observation of the fireline location is necessary), it represents a “leap” in capability, offering the ability to provide accurate predictions that, when coupled with the “holy grail of firefighting,” knowledge of firefighter and fire locations at all times, could provide a new era in firefighting strategy.
Current approaches to real-time data-driven modeling are based on operational-type models, i.e. simplified semi-analytic models that predict the propagation of a fire as a function of time. Operational models are different from Computational Fluid Dynamics (CFD) models, which are three-dimensional numerical flow solvers based on the Navier-Stokes equations and the basic principles used in fluid mechanics and heat transfer of conservation of mass, momentum and energy. While CFD models invariably can provide more accurate predictions because they provide a description of the coupling between the atmosphere and the fire, current operational models can run much faster than real time, which is essential for future practical applications. Operational approaches such as those deployed in FARSITE [1], PROMETHEUS [2] and PHOENIX RapidFire [3], adopt a front-tracking perspective. They may be physical models based on a simplification of complex processes; empirical models that rely on correlations to observed data; or semi-empirical models combining the two approaches. Almost all operational models are empirical or semi-empirical in nature, requiring adjustments from real observations to calibrate the model unknowns. CFD approaches combining fire spread model and a meso-scale atmospheric solver such as FOREFIRE/MesoNH [4] and WRF-SFIRE [5], allow simulations of wildfires that can make their own weather and are thus leading the state of the art in fire modeling.
Operationally-oriented fire spread models can be used to simulate fire growth using selected vegetation maps and wind-weather scenarios but offer no information on the probability of an area being impacted in the short term under multiple vegetation and weather scenarios. Ensemble-based modeling overcomes this limitation by generating hundreds of potential vegetation and weather scenarios, leading to the prediction of thousands of individual fires. The ensemble-based predictions result in probabilities for fire spread. Data-driven modeling then offers to take advantage of real-time data to reduce the scatter of the ensemble-based predictions around a small sample and to produce a more reliable forecast at future times. The FIREFLY system [5-6] jointly designed by CERFACS and the University of Maryland develops this idea using an Ensemble Kalman filter and is currently being evaluated against data from the Rim fire. The main quantity of interest – the rate of spread – is better assessed, providing a time-persistent correction in the simulated fire front and improving wildfire dynamics forecasting.
The challenge of real-time modeling, of course, is access to useful real-time data. Initial studies have all used the fireline location as observations; this necessitates remote sensing, particularly from airborne- or satellite-based sensors. These measurements detect fire location and may provide an estimate of the fire intensity for each pixel (fire radiative power, or FRP). While polar orbiting satellites such as Terra, Aqua, and S-NPP (with MODIS and VIIRS sensors, respectively), provide autonomous, synoptic observations of fire activity, both day and night, nominally twice a day from each sensor, this temporal resolution, and the corresponding spatial resolution, may not be adequate for real-time fire modeling. NOAA’s Geostationary Operational Environmental Satellite system (GOES) offers greater temporal resolution, but suffers in terms of spatial resolution. This applies to both post hoc model evaluation of a fire event or real-time predictions of fire spread. Therefore data fusion with various sources of remotely sensed data, as well as downscaling techniques, could improve remotely sensed data resolution to fill gaps.
Firelines with spatial resolution of approximately 10 m and temporal resolution of approximately 10 minutes are ultimately desired to achieve a reliable forecasting tool with accurate-enough predictions for fire dynamics. These requirements can theoretically be met with current satellite technology; however, these requirements may also be cost-prohibitive at the moment. Some of these problems could be alleviated with the deployment of unmanned aerial vehicles (UAVs) over a fire. However, the use of UAVs has separate jurisdictional issues which to date have limited their use for prescribed fires. NIROPS (USDA Forest Service National Infrared Operations) have shown that it is possible to capture firelines at good spatial resolution using an airborne infrared sensor. However, the low frequency of the fireline mapping (maps are made only once per night) is a limitation. Part of the problem is that the process is not automated. The use of drones, for instance the use of an MQ-1 Predator Remotely Piloted Aircraft (RPA) on the Rim fire in California, was successful in observing particular fires, but no permanent program has been established, most likely because of the high cost and UAV safety concerns.
Several changes in the near future may change this picture. Smaller and cheaper sensors, new satellites funded by private industry and advancements in sparsely networked data may provide new means for data to be captured from multiple sources and automatically compiled together. This could come from public and commercial satellites, equipped firefighting aircraft that already span a fireline and UAVs that are advancing in popularity and decreasing in cost. Obviously, without procedures for UAVs to deploy during a fire and relay that information in a timely manner to modelers, data-driven operational fire spread modeling may not be feasible. However, advancements in technology and policy are coming so quickly that we foresee that real-time fireline data will be available within a decade.
The technology necessary to provide real-time wildfire simulations is rapidly emerging. The fire research community should be prepared to utilize this new technology along with high-performance computing to systemically quantify uncertainties and improve model predictions. Currently, there is little large-scale effort to put the pieces into place. Software development, notably at NCAR (National Center for Atmospheric Research, USA), CERFACS (Centre Européen de Recherche et Formation Avancée en Calcul Scientifique, France), UAB (Universitat Autònoma de Barcelona, Spain), UMD (University of Maryland, USA) and UCSD (University of California, San Diego, USA), have provided many of the pieces required to assimilate wildfires. Investments in remote sensing suitable to fireline scales and/or geographical scales (i.e., with approximately 10-m spatial resolution and 10-min temporal resolution) and in cyberinfrastructure allowing real-time integration of fire spread models and sensor data are necessary to connect the dots. With support from funding agencies and the fire response community, we strongly believe real-time data-driven wildfire modeling could provide a paradigm shift in the way we design and manage fire emergency response to future fires.
Further references
[1] Finney, M. A. (1998) FARSITE: Fire Area Simulator – model development and evaluation. Forest Service, US Dept. of Agriculture, Research Paper RMRS-RP-4.
[2] Tymstra, C., Bryce, R.W., Wotton, B.M., Taylor, S.W., and Armitage, O.B. (2010) Development and structure of Prometheus: the Canadian Wildland Fire Growth Simulation Model, Canadian Forest Service, Information Report NOR-X-417.
[3] Chong, D., Tolhurst, K.G., Duff, T.J., and Cirulis, B. (2013) Sensitivity Analysis of PHOENIX RapidFire. Bushfire CRC, University of Melbourne.
[4] Filippi, J.-B., Pialat, X., and Clements, C.B. (2013) Assessment of ForeFire/Meso-NH for wildland fire/atmosphere coupled simulation of the FireFlux experiment. Proc. Combust. Inst., 34, 2633– 2640.
[5] Mandel, J., Beezley, J.D., and Kochanski, A.K. (2011) Coupled atmosphere-wildland fire modeling with WRF 3.3 and SFIRE 2011. Geosci. Model Dev., 4, 591–610, doi:10.5194/gmd-4-591- 2011.
[6] Rochoux M.C., Ricci S., Lucor D., Cuenot B. and Trouvé A. (2014) Towards predictive data-driven simulations of wildfire spread – Part I: Reduced-cost Ensemble Kalman Filter based on a Polynomial Chaos surrogate model for parameter estimation, Natural Hazards and Earth System Science 14:2951–2973. doi:10.5194/nhess-14-2951-2014.
[7] Rochoux M.C., Emery C., Ricci S., Cuenot B. and Trouvé A. (2015) Towards predictive data-driven simulations of wildfire spread – Part II: Ensemble Kalman Filter for the state estimation of a front-tracking simulator of wildfire spread, Natural Hazards and Earth System Science 15:1721–1739. doi:10.5194/nhess-15-1721-2015.
