Dynamic Multi-resolution Cellular Space Modeling for Forest Fire Simulation Xiaolin Hu Department of Computer Science Georgia Sate University, Atlanta, GA, 30303 xhu@cs.gsu.edu Lewis Ntaimo Department of Industrial and Systems Engineering Texas A&M University, College Station, TX, 77843 ntaimo@tamu.edu ABSTRACT Multi-resolution modeling is an important research topic in modeling and simulating complex systems. In this paper, the authors develop a multi-resolution discrete event cellular space model with multiple spatial resolution cells, i.e., the spatial sizes of different cells are different. The change of cells’ spatial resolutions happens dynamically and adaptively as the simulation proceeds. This work is developed in the context of forest fire simulation. Specifically, as the fire front spreads along the cellular space, cells close to the fire front change to higher resolution (smaller size) and cells far from the fire front change to lower resolution (larger size). The conceptual framework to support multi-resolution forest fire simulation is described. Implementation and preliminary results are presented and discussed. Keywords: Dynamic multi-resolution, Spatial resolution, Cellular space model, Forest fire simulation, DEVS 1. INTRODUCTION Multi-resolution modeling is an important research topic in modeling and simulating complex systems. As discussed in [Davis and Hillestad 1993], resolution is a multifaceted concept. A multi-resolution model might mean the model to have multiple temporal resolutions, or multiple spatial resolutions, or multiple levels of abstractions. This paper concerns models with multiple spatial resolutions while their “levels of abstraction” are the same. We develop a cellular space model with multiple spatial resolution cells, i.e., the spatial sizes of different cells are different. The change of cells’ spatial resolutions happens dynamically and adaptively as the simulation proceeds. We develop this work in the context of forest fire simulation. Specifically, as the fire front spreads along the cellular space, cells close to the fire front change to higher resolution (smaller spatial size) and cells far from the fire front change to lower resolution (larger spatial size). Fire spread is a complex natural propagation phenomenon that requires a great deal of computer storage space to accommodate the large-scale spatial and temporal data needed in modeling and simulating it. In cellular space fire spread models [e.g. Vasconcelos 1993; Wainer and Giambiasi 1998; Ameghino et al., 2001; Morais 2001; Ntaimo, et al., 2004] the forest is divided into small areas referred to as forest cells. Fuel and topographical conditions are generally assumed to be uniform across the forest cell. Spatial resolution deals with the resolution in the input spatial data required for fire spread simulation which include fuel type, elevation, slope and aspect. Raster resolutions of 25 to 50 meters are most commonly available for topographic and satellite data and seem to provide acceptable level of detail for heterogeneous landscapes [Finney 1998]. Therefore, forest cell resolution (size of the forest cell) affects the accuracy in representing the actual fuel and topographical conditions. Consequently, this cell resolution has influence on fire spread simulation results. Note that a high cell resolution, that is, smaller size cells, would represent the actual spatial conditions more accurately. However, dealing with high cell resolutions typically challenges efficient computer simulation. In [Barros and Mendes 1997], a Dynamic Structure Cellular Automata (DSCA) method was developed to represent only the active model cells instead of loading all the cells from the very beginning of a forest fire simulation. This paper takes a different approach where cells are initialized in a low resolution and then change to high resolution when becoming active. Achieving the “right” spatial resolution has been long research in computer modeling and simulation. Discrete time fire spread simulations such as FARSITE [Finney 1998] dynamically adjusts the simulation time-step to achieve a specified level of spatial detail determined by the distance resolution. Also, a process called rediscretizing [Richards 1990] is applied to achieve a required level of perimeter resolution, which is the maximum distance allowed between vertices of a polygon [Finney 1998]. The finest resolution used for the simulation must be dependent on the resolution of the spatial data grids used as input [Finney 1998]. Unlike the use of the cellular space approach for large scale high resolution environmental simulation, [Filippi and Bisgambiglia 2002] propose the use of a vector space. In this case a phenomenon is described by its dynamic shape and decomposed in several points that can move using a displacement vector. Each point is allowed to instantiate a new point if there is a change in the space properties or to obtain a better resolution model. Other related work includes Adaptive mesh refinement (AMR) [Berger and Oliger 1984; Berger and Colella, 1989], which refines the temporal and spatial resolutions for regions of