Definition of the forest fire hazard variables for the nearestneighbour forecasting method

摘要

The nearest neighbour method is a case based statistical approach used to predict the occurrence of natural hazards. In Switzerland it is applied to forecast both snow avalanche and forest fire danger at a local or regional level. This paper discusses the application of the nearest neighbour method to assess forest fire danger in the densely populated region of the Southern Swiss Alps (Canton Ticino). The method requires a database of forest fire occurrence and/or non-occurrence events for a particular region. Each event is parameterised using measured meteorological data (air temperature, wind speed, relative humidity, etc.), fuel bed data (composition and dryness) and behavioural data (weekday, holiday, land location). The data is used to construct an N-dimensional parameter space. The nearest neighbour method searches the parameter space to find “similar” days of forest fire activity. “Similar” is mathematically defined as the closest weighted distance in the selected parameter space. Different statistical approaches are used to weight the variables. This paper specifically discusses the choice of parameters, including parameters which account for human behaviour, and the weighting of the variables such that an optimal forecast is obtained. The different statistical approaches are compared. We present example forecasts in Southern Switzerland and show that newly developed genetic algorithms provide the best predictions of forest fire occurrence.