Mapping and Assessing Fire Vulnerability - Alberta & British Columbia: national parks region

Authors: Safwan Rahman; Thomson Lau; Brent Manza; Nicholas Marcheggiani


Mapping and Assessing Fire Vulnerability
Alberta & British Columbia: national parks region


Abstract

Vulnerability to fire in Banff National Forest

Assessing forest fire vulnerability would be ideal for fire hazard specialists so that they can focus on areas that are of highest importance, such that mitigation efforts may be thoroughly and efficiently applied. A well developed regression methodology showed that land cover was the significant factor that influences a regions vulnerability to fires. Moreover, temperature, slope and elevation accounted for about 30% of the significance in influencing fire risk. The aforementioned four variables were used as parameters in creating vulnerability maps in the Alberta, British Columbia national parks region. Finally a secondary study area located at Wabakimi Provincial Park in Northern Ontario, Canada was employed to evaluate our methodology’s credibility. The results of this showed consistency in our method.

Introduction



Forest fires are a growing concern around the world, which has negative human and ecological implications. Canada receives on average, 9100 fires each year; it is interesting to note that only 2% of fires account for 98% of the total burned area. While the majority of fires are small in magnitude, it currently costs Canada 300 million dollars each year to deal with the damages (Wagner, N.D).

Current research in forest fire vulnerability is subjective in determining a quality ranking system for factors affecting the risk of fire. This research project aims at identifying a quantitative approach in assessing fire vulnerability. Banff National Park in Alberta, Canada is used for this research.

The research approach leverages a regression analysis and Saaty's analytical hierarchy process to identify and justify the relation between physical factors in a forest and its vulnerability to fire.

Study Area


The study area chosen to conduct research on is Banff National Forest in Alberta, Canada. This study area was chosen due to the ease of available resources. Once the research was completed, the model created (described under Methodology section), was applied to Wabakimi National Forest in Ontario, Canada for further research and application.

Satellite image of Wabakimi Forest

Map of Banff showing historical occurrences of forest

fires by area of coverage (in hectares).

 

Data Sources

It is important to keep track of data sources for intellectual property and other organizational purposes. These data sources are provided to you so you may find the data as you desire for your own projects. Note that some sources are open source (free) and some are paid.

Variables

The following are the list of variables used to determine fire risk. The following factors (variables) carry some   weighted value in causing vulnerability to fire. 

1. Temperature
2. Slope
3. Elevation
4. Land cover type
1. Coniferous
2. Deciduous
3. Mixed wood
4. Shrubs

Methodology

There were three distinguishable sub-processes identified; the crux of the first was a regression analysis, the second involved a complete work-through of an analytical hierarchy process, and the third involved careful manipulation of a set of rasters and summing them up. The regression analysis served its purpose in indicating the level of importance of each variable contributing to fire risk. Saaty’s matrix played the role in the analytical hierarchy process, which ultimately formulated the final weights of each variable.

Finally, reclassifying and rescaling all input variable rasters were carried out using ESRI’s ArcGIS software so that they all were in a common factor to be summed up together to produce the final fire vulnerability map.

Regression Analysis


The regression analysis took each variable and plotted it against the size of the forest fire. When plotted against the size, we were able to determine how much a certain variable had an effect on the size of the fire. The graph below shows an example of the plotting.

OLS regression is used to specifically find parameters a and b, which is the regressed line’s intercept and slope, respectively. The line can take any number of input factors and plot against its x-axis, which in our case would be the fire determining variables (Statistics.com. 2010). Our study involved seven b parameters for the seven variables defined. Once data was collected from each variable, they were all plotted against the fire size and an OLS regression estimator was produced. The equation below shows the OLS regression equation.

The variables listed above include both topographic and land cover variables. The variable names are clearly listed and defined in the table below.

The following are the results from performing the regression analysis on all the variables. The table below the equation shows the coefficients while the equation below shows our OLS regression equation. 

From the table and equation, we can conclude that the land cover types play significantly higher weights towards the vulnerability of fire. Now since we wanted to use these coefficients as only indicators to test the weight of each variable, it was necessary to conduct a regression on only the temperature, slope, and elevation. 

The following were the results:

Analytical Hierarchy Process - Saaty's Matrix

An analytical hierarchy process (AHP) is a system for solving complex decision-making situations. In the case of our study, it will be used to decide the significance of each of the seven variables by applying certain weights to them. Each variable will be given a weight adding up to 100%. Thomas L. Saaty developed a method for applying the AHP and coming up with appropriate weights. His method involved setting up a matrix, which matched each variable with each other. Each variable was compared with each other and their level of importance was identified between them (Dyer, 1990, p. 249).

To perform Saaty’s AHP, we used an automatic matrix calculation generator to overcome time and difficulty constraints. The Canadian Conservation Institute provided us a free-to-use Saaty’s Matrix generator (Canadian Conservation Institute, 2005). We used the beta coefficients attained from the regression analysis to decide the level of importance of each variable to one another. 

The matrix decision making was formed in two separate steps. The first step evaluated the topographic data and the land cover as a single variable on its own. Therefore, the first matrix included an evaluation of only four variables. The goal from doing this was to figure out how accountable land cover is compared with the other three variables on a scale of 100%. The second step required us to evaluate only the four land cover variables against one another. 

Sample of AHP:

The results shows the final weights of each variable as determined by the automatic AHP calculator. Land cover accounts for about 70% of forest fire risk while temperature holds second highest, although significantly less. 

The following show the results for performing each land cover type separately. 

Raster Creation

The final process in our methodology required us to reclassify all our input rasters so they would be put into a common scale. All the rasters needed to be in a common scale in order to be summed together. If the rasters did not have the same scale, then all the values would ultimately be meaningless. For example, if we left the DEM showing elevation out of scale from slope, where slope has values ranging from 0-30 and DEM has 100 to 3500; the DEM value would be far overstated.

 According to our weights, elevation received the smallest weight; therefore the overstatement would show a significant inaccuracy in our final map product.

The input rasters for the topographic variables were reclassified to a scale of 0 to 10. We chose a scaling of 0 to 10 because a small range scale would make it easier for fire hazard specialists to whom our product is targeted to, to assess areas that need the most focus on. Regions with a value closer to 0 represent areas of low fire vulnerability and values closer to 10 show higher vulnerability. Table 8 shows how we reclassified slope, elevation and temperature in a 0 to 10 scale.

Reclassified raster values

The first part of the process defined land cover to have a weight of 70.01%, according to table 6. This means that all the values in table 7 had to be rescaled to be 70.01% of the four variable weighting scheme. Therefore, the 50% coniferous value had to be rescaled to 70%, since it was the land cover variable with the largest weight. This is done using a scaling factor derived by dividing . The common scaling factor is 1.357. Table 9 below shows the new rescaled values for land cover type.

Equation for final vulnerability map

Results

The figure to the right shows the final vulnerability raster map of Banff National Forest. Note that the highest vulnerability value is 9.5 and the lowest is 0.35. As indicated in figure n.n the most vulnerable regions are in predominant in the northern region. Most vulnerable regions are situated on coniferous land. This is in accordance with our beta values and our weighting scheme.

Below you can see two maps. One shows fire sites, and the other shows fire sites along with camp sites. You can interpret from the image below how close camp sites are indeed to fire presence locations. There is no definite correlation, but one can make assumptions (under their own risk).

Now with our newly created model, we apply it to Wabakimi National Forest.

It would however be tedious to continuously recreate all the aforementioned steps in finding the vulnerability of forest fires elsewhere. Hence we designed our own model shown below using ESRI ArcMap's model builder. In this model, as long as the desired rasters are entered into input sections, the result should give you a map, such as the one shown above. 

References

Akpinar, E., Usul, N. GIS in Forest Fires. ESRI. Retrieved 23 May 2010 <http://proceedings.esri.com/library/userconf/proc05/papers/pap1052.pdf>.

Analytical Hierarchy Process (AHP) Program. Canadian Conservation Institute. 13 May 2005. Retrieved on June 29

Barbosa, M.R., Seoane, J.C.S., Buratto, M.G., Dias, L.S.O, & Raival, J.P.C., Martins, F.L. (2009). Forest fire alert system: a geoweb gis prioritization model considering land susceptibility and hotspots-a case study in the carajas nation forest, Brazilian Amazon.. International journal of geographic information science, 24(6), 873-901.

Chuvieco, E, & Congalton, R.G. (1989). Application of remote sending and geographic information systems to forest fire hazard mapping. 29, 147-159.

Dyer, James S. Remarks on the Analytical Hierarchy Process. Management Science. Vol 36, No 3. EbscoHost. March 1990. <http://web.ebscohost.com/ehost/pdfviewer/pdfviewer?vid=2&hid=9&sid=4e27ed13-140d-4763-a8fe-bc2edeac60d9%40sessionmgr13>.

Flannifan, M.D., Stocks, B.J., & Wotton, B.M. (2000). Climate change and forest fires. The science of the total environment, 262, 221-229.

Jaiswal, R.K., Mukherjee, S, Raju, K.D., & Saxena, R. (2002). Forest fire risk zone mapping from satellite imagery and gis. International journal of applied earth observation and geoinformation, 4, 1-10.

Madry, Scott; Cole, Mathew L.; Siebel, Scott. Archaeological Predictive Modeling: Method and Theory. Retrieved 25 June 2010 < http://www.informatics.org/Poster_SEAC-05.pdf>. Ordinary Least Squares Regression. Statistics.com. 2010. Retrieved on 20 July 2010 <http://www.statistics.com/resources/glossary/o/olsregr.php>.

Wagner, C.V. (n.d.). The Canadian encyclopaedia: forest fire. Retrieved from http://www.thecanadianencyclopedia.com/index.cfm?PgNm=TCE&Params=a1ARTA000290