Determining map partitioning to minimize wind field uncertainty in forest fire propagation prediction
Introduction
Forest fires are significant natural hazards around the world, especially in places with hot, dry summer seasons such as Mediterranean countries, California or Australia. To fight these hazards and use the available resources in the best possible way, it is necessary to have an accurate prediction of their evolution beforehand. So, propagation models have been developed to determine the expected propagation of a forest fire [12], [2], [9]. Such propagation models require several input parameters representing the scenario where the fire is taking place to produce the predictions of the propagation. These parameters include the digital elevation map, the vegetation map, the vegetation features, and meteorological conditions, among others. Some parameters are well known, but the values of other input parameters are obtained or estimated from indirect measurements. Such indirect estimations imply a degree of uncertainty concerning the values of the parameters that provoke uncertainty in forest fire propagation prediction.
The parameters that most significantly affect forest fire propagation are wind speed and direction [1]. These parameters can be measured at meteorological stations or estimated from meteorological models, but, in both cases, they are obtained at a very low resolution, typically some a distance of kilometers. This fact is critical because, as is well known, the meteorological wind is modified by the topography of the terrain and, therefore, the values of the wind speed and direction at one point of the terrain are different from the values at other points, depending on the terrain topography, and the values at low resolution are not representative of the actual situation. This may imply that the predictions provided by forest fire simulators are not feasible and not very useful in real operation.
To estimate the wind speed and direction at each point of the terrain, it is necessary to apply a wind field model that determines those values at each point while taking the terrain topography into account. Then, the wind field generated by the wind field simulator is used as input of the forest fire propagation simulator, coupling both simulators, wind field and forest fire, to improve the accuracy of forest fire propagation predictions [4].
Moreover, meteorological wind varies dynamically over time. So, when predicting fire propagation, it is necessary to consider the meteorological forecast provided by meteorological centres at different time steps and calculate the wind field for each time step. During the forest fire simulation, the corresponding wind field must be applied at each time step. So, if the meteorological forecast provides a wind value each hour and the forest fire propagation is predicted for the next 6 h, it is necessary to calculate 6 wind fields (one for each hour) and introduce them into the forest fire simulator.
FARSITE [5] is a widely used forest fire propagation simulator that has been extensively tested on real fires and produces successful results. It is a fire behavior and fire growth simulator that incorporates both spatial and temporal information on topography, fuels and weather. It includes temporal variation in fire conditions. FARSITE is an elliptical wave propagation simulator and avoids a typical problem in cell-based simulators of reproducing the fire shape in heterogeneous conditions, due to their reduced number of propagation paths.
For the wind field simulator, WindNinja [7] is a mass conservation wind field simulator that, given a meteorological wind and a digital elevation map, generates the wind field at the needed resolution (30 m). One of the main advantages of WindNinja is that the output generated can be directly used as input by FARSITE. This simplifies the coupling of both models. Fig. 1 shows the coupling of WindNinja and FARSITE.
WindNinja has also been used in other simulation environments and coupled to simulators such as WildFireAnalyst [11], Phoenix RapidFire [15] and WIFIRE [3].
However, when the studied terrain map is large (for example, 45 km × 45 km) and uses a high resolution (for example, 30 m), coupling a wind field model with the forest fire propagation implies a significant increase in the execution time that is not affordable since the propagation prediction has strict real time constraints in order to be operational. Therefore, it is necessary to apply computational methods to reduce the execution time of both, the forest fire propagation model and the wind field model. In this work, the main goal is to reduce the WindNinja execution time for large terrain maps such as 45 km × 45 km and 1500 × 1500 cells, in order to make it operational in real scenarios while accomplishing operational real time constraints.
A map partitioning approach to parallelize WindNinja was proposed in previous works [14]. In this approach, the terrain map was divided into smaller parts, and the wind field can be calculated in parallel on each part of the map and the wind fields of the different parts are then aggregated to form the global wind field. As stated, the map partitioning strategy can reduce WindNinja execution time, but it can also introduce a large error in wind field that can make the approach unfeasible. On the one hand, WindNinja presents certain effects in the cells close to the map border that make that the values of the wind parameters in those cells are unreliable. Therefore, it is necessary to introduce a certain degree of overlapping among the different parts to avoid such limitations. On the other hand, when the map is partitioned, the system of equations is solved for each part, but, in this way, there is a loss of global information that modifies the solution and the consequent wind field. These facts have been brought to light by preliminary studies, and it has been proved that it is necessary to develop a complete methodology to reach a tradeoff between execution time and wind field accuracy. The result must be a map partitioning methodology that reduces execution time below operational requirements and provides accurate partitioned wind field that, when gathered, reaches insignificant fire propagation differences.
In this methodology several factors such as map part shape, map part size, map resolution, part overlapping, and others must be considered. Some preliminary studies have been carried out [13], and it has been proved that some of these parameters are interrelated. The goal of this work is to establish a complete methodology that, given a map size, an operational wind field computation time and a maximum forest fire propagation error, is able to provide the map partitioning required to accomplish such requirements or to indicate the values that cannot be reached.
In this paper, such a complete map partitioning methodology to accelerate wind field calculation is presented. So, the paper is organized as follows. Section 2 describes the main features of WindNinja and presents its main limiting factors. Section 3 introduces the map partitioning approach to overcome the execution time and memory limitations of WindNinja. Section 4 presents the methodology to determine the most adequate map partitioning to minimize wind field uncertainty and reduce execution time. Section 5 summarizes the results of the experimental study carried out. Finally, Section 6 shows the main conclusion of this work.
Section snippets
WindNinja wind field simulator
Wind speed and direction are two of the parameters that most significantly affect fire propagation. Such parameters suffer from temporal and spatial variation. Such variations introduce a high degree of uncertainty in the results of forest fire propagation prediction. As is well known, wind is not constant. Therefore, a weather forecast model must be introduced to predict the future evolution of the wind. On the other hand, the meteorological wind is modified by the topography and vegetation
WindNinja parallelization using map partitioning
The approach to overcoming the WindNinja limitations is to use map partitioning. In this way, a global map is partitioned into a certain number of parts and the wind field is calculated simultaneously for each part. In this way, the execution time is reduced in proportion to the number of parts, and the amount of memory required is also reduced. This scheme has been parallelized using a Master/Worker paradigm implemented in MPI where the Master process carries out the map partitioning and
A complete methodology to determine map partitioning
The main constraint of forest fire propagation prediction is execution time. Therefore, it is a crucial point that must be taken into account as the main priority. So, the first studies are directed to guarantee the execution time of WindNinja. But, forest fire propagation accuracy is also a key point, and it is necessary to estimate the difference between the forest fire propagation prediction obtained with WindNinja map partitioning from the propagation obtained when WindNinja is executed on
Experimental results
The first experiment was carried out on a terrain map corresponding to La Jonquera (Spain) in a place where there was a fire in July 2012. The map area is approximately 50 km × 50 km. The execution time required to calculate this complete map with a 30 m resolution is 1, 402 s. Considering a maximum time of 100 s, Eq. (9) provides λ(t) to be 206, 533. The map resolution is determined to be 30 m by applying Eq. (12). The overlapping obtained from Eq. (13) is 50, and the number of parts from Eq. (14) is
Conclusion
Wind speed and direction are parameters that affect forest fire propagation dramatically. So, an accurate estimation of such parameters is crucial in predicting the fire propagation precisely. However, meteorological wind is modified by terrain topography, and a different value of wind speed and direction is effectively found at each point of the terrain. To overcome such uncertainty in wind parameters, it is necessary to introduce a wind field simulator and couple this simulator with a forest
Acknowledgements
This research has been supported by the Ministerio de Ciencia e Innovación under contract TIN-2011-28689-C02-01, by Ministerio de Economía y Competitividad under contract TIN2014-53234-C2-1-R and partly supported by the European Union FEDER (CAPAP-H5 network TIN2014-53522-REDT).
Gemma Sanjuan received a Bachelor degree on Physics for UAB in 2010. She also got the Mechanical Engineering degree from UPC in 2011. Afterwards she received the M.Sc. degree on Modelling for Science and Engineering in 2012 and the M.Sc. degree on High Performance Computing, Information theory and Security in 2013, both form UAB. In October 2013 she joined the High Performance Computing Applications for Science and Engineering group at the Computer Architecture and Operating Systems department
References (15)
- et al.
Enhancing wildland fire prediction on cluster systems applying evolutionary optimization techniques
Future Gener. Comput. Syst.
(2005) - et al.
Towards an integrated cyberinfrastructure for scalable data-driven monitoring, dynamic prediction and resilience of wildfires
Procedia Comput. Sci.
(2015) - et al.
Coupling wind dynamics into a DDDAS forest fire propagation prediction system
Procedia Comput. Sci.
(2012) - et al.
Wind field uncertainty in forest fire propagation prediction
Procedia Comput. Sci.
(2014) Estimating Wildfire Behavior and Effects. General Technical Report INT. Dept. of Agriculture, Forest Service
(1976)Farsite: Fire Area Simulator: Model Development and Evaluation
(2004)- et al.
Simulating diurnally driven slope winds with WindNinja
Cited by (10)
Using efficient parallelization in Graphic Processing Units to parameterize stochastic fire propagation models
2018, Journal of Computational ScienceCitation Excerpt :Fuel and air humidity change because of fire. Fuel type is usually considered as the average of typical vegetation of the area, while weather and topography are generally interpolated values that can be calculated for example using WindNinja [2,5,11,4]. The cellular automaton (CA) approach was efficiently used by Russo et al. [12] to run thousands of simulations to build hazard maps that have been shown to be robust and efficient in predicting fire spreading behavior in several cases.
Numerical study of fire spread using the level-set method with large eddy simulation incorporating detailed chemical kinetics gas-phase combustion model
2018, Journal of Computational ScienceCitation Excerpt :For example, Bianchini et al. [13] developed an Evolutionary-Statistical System (ESS) that attempt to adjust input parameters in real time. Sanjuan et al. [14] developed a map partitioning method to minimise wind field uncertainty for fire spread prediction. Application of mathematical modelling on bushfires has become increasingly popular due to the rapid advancement of numerical methodologies and computational power.
Hybrid application to accelerate wind field calculation
2016, Journal of Computational ScienceRoute to exascale: Novel mathematical methods, scalable algorithms and Computational Science skills
2016, Journal of Computational ScienceA hybrid intelligence system based on relevance vector machines and imperialist competitive optimization for modelling forest fire danger using GIS
2020, Journal of Environmental Informatics
Gemma Sanjuan received a Bachelor degree on Physics for UAB in 2010. She also got the Mechanical Engineering degree from UPC in 2011. Afterwards she received the M.Sc. degree on Modelling for Science and Engineering in 2012 and the M.Sc. degree on High Performance Computing, Information theory and Security in 2013, both form UAB. In October 2013 she joined the High Performance Computing Applications for Science and Engineering group at the Computer Architecture and Operating Systems department of UAB. She enrolled the Computer Science Ph.D. programme and she is currently developing her PhD thesis on acceleration of wind field models calculation by applying data and functional parallelism advised by Prof. Tomàs Margalef.
Carlos Brun received the degree of Computer Science Engineer from Universitat Autònoma de Barcelona in 2008. He also received the M.Sc. degree on High Performance Computing from the same university and finally in 2014 he received the Ph.D. also in High Performance Computing. His research has focused on the coupling of complementary models to forest fire propagation models.
Tomàs Margalef got a BS degree in physics in 1988 from Universitat Autònoma de Barcelona (UAB). In 1990 he obtained the M.Sc. in Computer Science and in 1993 the Ph.D. in Computer Science from UAB. Since 1988 he has been working in several aspects related to parallel and distributed computing. Currently, his research interests focuses on development of high performance applications, automatic performance analysis and dynamic performance tuning. Since 1997 he has been working on exploiting parallel/distributed processing to accelerate and improve the prediction of forest fire propagation. Since 2007 he is Full Professor at the Computer Architecture and Operating systems department. He is an ACM member.
Ana Cortés received both her first degree and her Ph.D. in Computer Science from the Universitat Autònoma de Barcelona (UAB), Spain, in 1990 and 2000, respectively. She is currently associate professor of Computer Science at the UAB, where she is a member of the High performance Computing Applications for Science and Engineering of the Computer Architecture and Operating Systems Department. Her current research interests concern performance engineering of high performance computing environmental sciences applications.