WILDLAND SURFACE FIRE BEHAVIOUR: A SPATIAL SIMULATION MODEL FOR OPERATIONAL EMERGENCY MANAGEMENT

Wildfires affect vegetation dynamics, geomorphological processes, biogeochemical cycles, atmospheric chemistry, and climate, posing a severe threat to human lives and activities interacting with the natural system. As both fire activity and wildland-urban interface exposure are expected to increase under future climate projections, the improvement of our ability to promptly predict wildland fire behaviour, in terms of expected intensity and geographic patterns, has become a tangible need. The general purpose of this research is to investigate on wildland surface fire behaviour simulation models and to support disaster managers in optimising decision making processes in wildfire risk management in a Mediterranean-type climate region, namely Sardinia, Italy. This project is intended to pursue two major objectives: (i) develop and validate a predictive spatially distributed wildland surface fire behaviour simulation model intended for operational use; (ii) design and implement a geospatial decision support system to provide decision makers with appropriate strategies and tools for an integrated wildland fire risk management. Predicting wildland surface fire behaviour requires a deep understanding of the influence of environmental parameters that act as drivers of the fire spread, including geomorphometrical variables, meteorological conditions, and fuel characteristics, on fire descriptors, such as the rate and direction of the maximum fire spread, the eccentricity of the ellipse approximating the fire shape, the intensity of the fire front, and the flame length. The Rothermel’s mathematical model for predicting surface fire spread in wildland fuels is currently the most extensively used method to estimate fire descriptors, especially for operational purposes. The application of the Rothermel’s model for simulating the behaviour of ongoing wildfires calls for the need of a technique for continuous monitoring of the spatiotemporal variability of weather conditions and fuel characteristics, such as fuel height, loading, and moisture content, in the pre-fire environment. Firstly, freely available data sources and remote sensing products and datasets have been investigated to define a pre-processing methodology for the near real-time estimation of the drivers of fire spread. Secondly, the need for flexibility in handling the equations of the Rothermel’s and associated models, together with the necessity of integrating corrections and updates, have led to an original implementation of a computer algorithm that evaluates the fire descriptors as defined by the extended Rothermel’s mathematical model. Then, a proxy model of this implementation has been developed using a machine learning ensemble method in order to analyse the interdependence of the drivers and to understand their relative importance in predicting fire descriptors. Furthermore, the proxy model for predicting fire spread across heterogeneous landscapes has been integrated into an agent-based simulation model developed to predict the surface fire behaviour and growth with the aim of providing fire management authorities with timely information on the expected progress of the fire front. Finally, the developed simulation model has been applied to and validated against historical wildfire events recorded in Sardinia, Italy, to evaluate its performance in terms of predictive capacity. The effects of fire suppression activities have also been simulated according to the availability of accurate information on timing and location of interventions that effectively extinguished the fire’s spread. As a whole, the developed wildland surface fire behaviour simulation model, together with the pre-processing methodology, have resulted in a satisfying accuracy in terms of quantitative agreement between modelled and observed patterns of fire growth. The adoption of the proxy model instead of its original implementation has guaranteed a significative reduction of the computing time in the face of a limited loss in accuracy at the scale of the analysis if compared with the original implementation of the Rothermel’s equations. Results of the validation suggest the model’s suitability for operational uses for predicting wildland surface fire behaviour. The predictive ability of the simulation model could reasonably benefit from the inclusion of some additional mathematical models simulating the potential evolution of the surface fire towards passive or active crown fires or spotting fires. Moreover, major improvements could be granted by implementing in the agent-based simulation model a wider range of fire suppression activities and techniques, ranging from ground to aerial interventions. The proposed predictive model could become a valid tool for the optimization of risk planning, prevention, and management activities. Within the context of this project, three modules of a geospatial decision support system have been designed and implemented with the aim of improving the efficiency of risk management strategies and reducing expected impacts and potential damage. The first module is a dynamic workflow of actions and represents the core of the decision support system. This module aims to guide decision makers in carrying out the procedures of the intervention model compliant with the legislative framework. The workflow is then supported by a second module, a customised version of a geographic information system with dynamic forms designed to support users with limited expertise in geodatabase management. This module will incorporate a structured relational geodatabase storing (i) scenarios of wildfire events, produced by means of the developed predictive wildland surface fire behaviour simulation model, (ii) existing institutional wildfire susceptibility, hazard, and risk maps, (iii) available resources and socioeconomic exposed values, and (iv) real-time data from field surveys. Finally, the decision support system will provide authorities and technicians with a third module composed by web applications for mobile field data collection and sharing. This research strived to investigate principles and accepted theories on the complex dynamics of wildland surface fire behaviour and to shed light on the need for a better understanding of the difference between real and simulated fire behaviour in terms of the importance of the drivers of fire spread in predicting fire spread and growth. The project also tried proposing solutions integrating remote sensing and machine learning techniques with the aim of improving the applicability of near real-time simulation models as well as the effectiveness of decision-making strategies.