Meso-scale wildfire modeling using artificial
neural networks.
Ronald Joseph McCormick - 2001
[UMI Proquest
Full
Citation]
Modeling wildfire spread patterns is a complex problem
involving long-term fuel accumulation (site history) with short-term thermodynamics.
Two general approaches to modeling wildfire spread patterns are fine-scale
mechanistic or broad-scale probabilistic. Mechanistic approaches scale locally
(micro-theory) to what keeps a fire burning while fire spread in probabilistic
models is constrained by the rate of percolation across the landscape (macro-theory).
Changing spatial and temporal scales of fire environment variables lead to
the inherent unpredictability found in middle number systems. Extant fire
models lose predictive power when subtle shifts in environmental variables
cause a qualitative change in fire behavior. Artificial neural networks (ANNs)
are designed for problems with cross-scale relationships that produce non-linear
changes in system behavior (meso-theory). Even though the system appears
middle number, the ANN recasts system structure until, at an appropriate
level of analysis, prediction becomes possible. The difficulty with ecological
systems is they invite being cast as complex, and complex systems require
different causal models. A systems approach incorporates the explanatory
power of positive and negative feedbacks and the recognition of emergent
system behavior. Because complex systems do not invite definitive answers,
we need complex systems methodologies like ANNs to offer prediction with
good explanatory power. An ANN wildfire spread model was developed that integrates
across scales of fire environment variables. Preliminary results support
the proposed meso-theoretical fire environment definition and ANN-based modeling
approach.