Inverse modelling in wildfire spread forecasting: towards a data-driven system
2018
Wildfires are an ecological phenomenon inherent to earth dynamics and widely spread over the globe. In addition to the environmental impact, when wildfires grow beyond controllable magnitudes, they pose a principal threat to human lives and properties. On many countries, the rural abandonment of last decades, the forest continuity growth and the Wildland Urban Interface increase are exposing entire communities and rendering them vulnerable to a major fire event. Coupled together with a global warming that seems to be enlarging and worsening wildfire-prone weather conditions, the wildfire problem is becoming a recurrent and repetitive natural hazard that is in urgent needs of research development, planning and organizational changes to minimize its impact. In this context, the thesis at hand focuses on the development, implementation and initial validation of a wildfire perimeter spread prediction model that might help emergency responders on taking sound decisions to efficiently employ resources and protect valuable assets. This forecasting model is a particular implementation of a data-driven system. That is, available data are used to improve and calibrate the spread model results with the aim of delivering a more accurate and timely forecast of the fire spread for the upcoming hours. This thesis builds up the mentioned system by increasing its complexity and tackling required improvements and adaptations on fuel characterization and wind projection on topography. Initially, a simplified proof of concept that uses front perimeter (isochrones) evolution extracted from infrared imagery of the fire is challenged with data from real-scale burning experiments conducted in Australia. Despite the positive outcome of this initial investigation, some advancements are identified to further upgrade the system. Thus, the following chapters focus on the fuel and wind sub-models together with the spread model topographic upgrade and the different mathematical algorithms and strategies necessary to conduct the data-driven process. Regarding fuels, the thesis presents an in-depth analysis of fuel characterization to be used by fire spread models. This is done by a thorough sensitivity analysis of the most commonly used fuel characterization systems. In the light of these results, a simplified model that integrates all different fuel properties is proposed to be used by the data-driven framework at hand. To properly resolve the wind interaction with the terrain and to couple it into the data-driven system, the WindNinja diagnose software is employed. However, long computational times do not allow for its integration into any data-assimilation strategy. Thus, a full interpolating framework is developed and validated to allow fast and computationally inexpensive wind field updates. This key element becomes then a cornerstone of the full data-driven approach. For the optimization process (embedded into any data-driven systems) six different mathematical algorithms were compared and evaluated. Three of them being line-search strategies and the other three being global search. It was found that the algorithm selection has an impact on the final results in terms of forecast accuracy and computing time. Finally, the overall system is verified and validated using two source of available data: (1) well characterized, homogeneous slope, medium-scale experimental fires conducted in Portugal and (2) with synthetically generated fronts reproducing a real large-scale fire. These validations were aimed at studying the overall performance, checking the system functionality and highlighting possible flaws and necessary improvements if the tool is to be deployed in a real emergency situation. Whereas the results show the potential of the approach by delivering an acceptable forecast usable for emergency responders, further validations are required to check the robustness and reliability of the tool before using it in operational situations.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
0
References
0
Citations
NaN
KQI