| Summary: | A framework is presented for a numerical prediction scheme with a prediction accuracy of
greater than 70% at Bear Pass, British Columbia, Canada. One way analysis of variance and
canonical discriminant analysis are used to identify the principal variables that allow
discrimination between avalanche and non-avalanche time periods. The optimum variable set
is then used in a Mahalanobis-metric nearest neighbours model to determine the observations
in the historical database that most closely represent conditions of the forecast period. The
information about these neighbours can be used by the forecaster to aid in making an
avalanche forecast.
In this study, a univariate analysis of the meteorological and snowpack data from The Pass is
used to assess the contribution of each of the variables for use in avalanche prediction. Using
these data, the snow-climate of the area is assessed and it is shown that Bear Pass may be
classified as a maritime snow climate, although some elements of a transitional snow-climate
are evident.
A canonical discriminant analysis is performed for both a two-group discrimination into
avalanche and non-avalanche periods and for a three-group discrimination into wet, dry and
non-avalanche periods. For the three-group discrimination, a two-stage discrimination is
preferred, involving first discriminating into wet-avalanche or dry-avalanche periods and
then discriminating between wet or dry-avalanche and non-avalanche periods. The analysis is
performed for all areas in a combined analysis and also for individual sub-areas defined
within The Pass. Improvements on classification rates to three out of the four sub-areas are
found compared to the analysis for the whole Pass. Various parameters that affect the nearest neighbour model performance are analyzed. The
effect of group size distribution is assessed, as well as the number and threshold number of
neighbours that form the basis for the forecast decision. The effect of dimensionality on the
model is also analyzed. The Mahalanobis-metric nearest neighbours model is compared to
Cornice, a nearest neighbours model developed for avalanche forecasting in Scotland.
Classification rates for the two models are found to be very similar.
Finally, preliminary findings from a new method using time patterns of predictor variables is
presented. The method makes use of time series data from remote weather stations. Sections
of the time series are similar to present conditions are identified through a deconvolution
process. It is hoped that this method may offer improvements to the conventional nearest
neighbours method by implicitly assuming that meteorological conditions are correlated in
time.
|
|---|
