Switzerland: Fire Danger Prediction in the South of Switzerland (IFFN No. 16)
Fire Danger Prediction in the Southern Part of Switzerland
(IFFN No. 16 – January 1997, p. 2-6)
Introduction
The general increase in the occurrence of forest fires since the sixties (Conedera et al. 1996) makes it more and more necessary to improve the forest fire management methods in southern Switzerland. The development of Fire Danger Prediction methods is considered to be one of the most important elements of an effective forest fire management strategy.
In the southern part of Switzerland humans have a very strong direct and indirect influence on fire occurrence. On the one hand nearly all forest fires are human-made, either through carelessness or intention. On the other hand fire brigades in southern Switzerland are very well organised and trained in fire fighting. With the support of fire prediction systems, which can give rational information about the temporal and local evolution of forest fire risk, we can profit very much.
Studies of forest fire prediction in southern Switzerland started in the beginning of the nineties and were carried on within the scope of the Swiss National Research Program 31 (NRP 31) “Climate Change and Natural Disasters”. Finally two modules were developed in the frame of the EU research project Minerve II:
- a statistical model based on the Poisson distribution and
- a hybrid expert system for the spacial prediction of wildfire danger.
These two modules are totally complementary and will be used in one fire risk prediction package.
In this second contribution on forest fire research in Switzerland we will report on the theoretical aspects of these two different but complementary approaches and the operational use of this fire risk prediction package.
Statistical Poisson Model
In frame of the EU project MINERVE II we have developed a general statistical methodology for the prediction of forest fires in the context of Poisson models. Quantitative and qualitative tools are given for the assessment of different models, and some theoretical decision considerations are also discussed for the practical application of fire danger prediction methods in general. Case studies from France, Italy, Portugal and Switzerland illustrate that Poisson models incorporating a fire danger index with other highly important explanatory variables are always superior to the empirical use of a single fire hazard index.
Summarizing our experience with the Swiss, French, Italian and Portuguese data the following conclusions can be drawn:
- The Poisson framework is simple and yet flexible enough to allow for a statistically sound modelization of the occurrence of forest fires. It can be approximated by a logistic regression – a procedure available on most statistical software packages. As compared with other danger indexes, it has a clear probabilistic interpretation, which is a definite advantage in risk assessment.
- Efficient prediction of the probability of forest fire occurrence requires:
-
- indicator variables coding subregions of the domain of investigation;
- indicator variables coding socio-economic factors, such as weekday versus holiday, legislation etc.;
- indicator variables coding special meteorological events with threshold behavior (e.g. Föhn);
- indicator variable coding seasonal effects. It is often simple and better to perform separate analyses for different seasons;
- Dryness index: within a meaningful model incorporating many explanatory variables the choice of a dryness index is, up to a certain point, a matter of taste and simplicity. The Canadian index, or sub-index thereof, IP, IREPI and the new Swiss index ETP (ET) are often in the top group;
- The variables describing the past history of fire events seem to always play a very important role;
- Synthetic danger indexes (i.e. including two or more danger indexes, possibly with interactions) seldom improve performance.
- The goodness of fit is qualitatively acceptable, especially when grouping breaks down the calendar structure of the data. The models tend to underestimate but still recognize extreme peak values
- In the absence of realistic cost functions it appears reasonable to choose the decision rule, which predicts, in the long run, the same number of events as there are observed events. The decision based upon the resulting cut-off point for the predicted probability of the event, as obtained from the Poisson model, leads in all instances to the highest overall rate of correct alarm. The improvement w.r.t. to the corresponding rules based on the danger indexes alone is substantial and increases when the average probability of the event decreases.
Spatial prediction
The method we used to perform spatial predictions of wildfire consists in merging case-based reasoning and knowledge-based reasoning (AI method) with the goal to establish probabilities of fire occurrence at various locations in a region.
The main steps of the prediction model development were:
Step 1: Terrain tessellation
The region (Sottoceneri in our study) was divided into homogeneous panels according to slope angles, fuel type and aspects (Fig. 1). This work was done by experts from FNP-SSdA (forest engineers).
Fig.1. Terrain tessellation
Step 2: Knowledge representation
Observations were organized into three types of information: meteorological data, terrain data (including topography and fuel descriptions) and past fire data. Experts’ knowledge was translated into production rules.
Step 3: Diagnostic modeling
An expert system was design to perform probabilities of fire ignition for each panel defined at step 1 (Fig.2). This expert system uses meteorological observations of the day to be analysed (measurements taken at a reference point) as well as the fires which happened during similar days in the past. A specific nearest neighbours procedure is called by the model to determine what were similar days.
Fig.2. Diagnostic process
Step 4: Implementation
The model was implemented to run on a PC. A user-friendly interface makes it usable by practitioners who need efficient tools. The software very quickly delivers a diagnostic showing the most dangerous panels (Fig.3).
Step 5: Evaluation
The model was tested on a very large sample: 362,604 diagnostics were compared to observed events. The results of the test demonstrate that the model is very reliable when locating the most dangerous panels: this model is a good classifier performing reliable spatial predictions. So it may be used as a complement to the Poisson model which gives good temporal diagnostics.
Fig.3. Diagnostic screen
This study was a first attempt to solve the problem of wildfire forecasting by using Artificial Intelligence methods. The first results make us think that we should go deeper into this approach in order to improve the diagnostic process as well as the knowledge base.
Application of the Fire Risk Prediction Package
During the ongoing fire season in southern Switzerland (December to April) the statistical model is installed at the branch station of the Swiss Meteorological Service in the southern part of Switzerland in Locarno-Monti. The statistical model is tested there in operational conditions. For the expert system Pyrolog further development is needed, but we think that it will be possible to test it in operational conditions soon. The link between the two models can then be made.
Conclusion
Rational Forest Fire Danger Prediction is an essential part of a modern forest fire management system. In our research programs we have tried to develop a powerful method to support short term fire prevention and fire fighting decisions for southern Switzerland. We think that the proposed prediction package will be a useful tool for the Weather Forecast Service and the responsible authorities of fire management and fire-brigades in their fire prevention and fire fighting activities.
References
Conedera, M., P. Marxer, C. Hofmann, W. Tinner, B. Ammann. 1996. Forest fire research in Switzerland. Part 1: Fire ecology and history research in the southern part of Switzerland. Int. Forest Fire News No. 15, 13-21.
From: Marco Conedera Daniel Mandallaz FNP Sottostazione Sud delle Alpi D-WAHO casella postale 2014 ETH Zürich CH – 6501 Bellinzona CH – 8092 Zürich
Fax: +41-91-821-5565
Fax: +41-1-632-1127 Tel: +41-91-821-5562 Tel: +41-1-632-3186 e-mail: conedera@wsl.ch e-mail: mandallaz@pfe.waho.ethz.ch
Robert Bolognesi
Paolo Ambrosetti EISLF ISM Osservatorio Locarno-Monti Flüelastr. 11 via Mit della Trinitä 146 CH – 7260 Davos CH – 6605 Locarno-Monti
Fax: +41-81-417-0110
Fax: +41-91-756-2310 Tel: +41-81-417-0153 Tel: +41-91-753-2316 e-mail: bolo@slf.ch e-mail: pam@otl.sma.ch
