Abstract

An ad hoc method for forecasting future fire/weather activity is to scan previous
years' weather patterns and identify the one with the closest match to the current one.
This presentation describes methods for formalizing this approach using time series
analysis and clustering methods. One approach is to fit state space models to the
data, then to use Kalman filters and an EM algorithm to do the clustering and
prediction. Another method involves the use of log-Gaussian Cox processes. The
techniques are compared by simulation and illustrated on data from British Columbia
and Alberta, Canada.