Canadian Wildland Fires Story

Overview

Canada has about 9% of the world’s forests. Each year over the last 25 years, about 7,300 forest fires have occurred. The total area burned varies widely from year to year but averages about 2.5 million hectares annually. Some are uncontrolled wildfires started by lightning or human carelessness. A small number are prescribed fires set by authorized forest managers to mimic natural fire processes that renew and maintain healthy ecosystems Wildland fires is a challenge for forest management with the potential to be at once harmful and beneficial. Forest agencies work to harness the force of natural fire to take advantage of its ecological benefits while at the same time limiting its potential damage and costs. This makes fire control strategies a vital component of forest management and emergency management in Canada.

Objective

Analyzing previous wildland fire data among the Canadian provinces and national parks to optimize the fire control strategies.

Context

Wildfires is a project that I built as part of my data immersion course at CareerFoundry to demonstrate my mastery of data analysis. I as a data analyst, helped CFS (Canadian Forest Service) by analyzing fire characteristics so they can prioritize and improve timing, locations, etc.

  • Jupyter Notebook
  • Tableau
  • Python packages and libraries:
    • Numpy
    • Pandas
    • Matplotlib
    • Skilearn
    • Statmodels
    • Folium
  • Data wrangling
  • Visualizations
  • Advanced analysis in Python including:
    • Geospatial analysis
    • Supervised ML (Regression)
    • Unsupervised ML (Clustering)
    • Time series analysis

Questions

  1. How fire ignites?
  2. What is the proportion of each cause?
  3. How is the wildland fire behavior in different provinces?

Solution

These three questions are solved by analysis of “cause” variable in the dataset.

“Lightening” and “Human-caused” have burned the largest sizes in the past 25 years. We prioritized “human-caused” fires (as a cause which can be controlled) and focused on weather conditions for further analysis.

The top- three provinces with the highest number of fires in the past 25 years are selected for further exploration: British Columbia, Alberta, and Ontario.

Stacked bar chart is used to present the proportion of each cause in each Canadian province.

Questions

Solution

In order to answer this question, we categorized the fires based on “size classes”.

Distribution of fires in Canadian provinces are presented in size classes dashboard.

The stacked bar chart shows that most of the fires in Canadian provinces occurred in small sizes which fall into Class A and Class B.

Northwest Territories have the highest proportion of Class E (largest size) fires with around 32.9% of fires which happened in this area in the past 20 years.

Questions

Solution

To find an answer to this question we looked for daily weather information for fire locations in Canada. The only open dataset that we were able to find with longitude and latitude was for Alberta, which happens to be among the top three Canadian provinces with the highest number of fires in the past 20 years.

We used the info of the closest weather station to a specific fire in Alberta to map the weather condition around the fire location.

Maximum Wind gust and Maximum temperature are the variables which we selected to conduct the analysis on.

Correlation Matrix was created in Python to determine the relationship between the size of fire and weather characteristics. A weak positive relationship was revealed between the size of fire and the maximum wind gust around the fire location in Alberta.

Questions

Observation

After analyzing and exploring data with regression, we found out there is a weak positive relationship between fire spread and wind gusts in Alberta over the last 20 years.

In the Scatterplot situated underneath, most of the larger fires (light and dark purple dots, which correspond to fire sizes averaging around 88 and 135 hectares, respectively) happen with higher maximum wind gust values. Smaller max wind gust values are associated with smaller fires (light and dark pink dots, corresponding to fire sizes averaging around 1 and 12 hectares, respectively).

This insight can be observed after applying clustering to the data and we were not able to find this relationship using supervised ML (linear regression).

Questions

Observation

We conducted time-series analysis in Python on historical report of Canadian wildland fires. To forecast the number of fires for upcoming years we applied ARIMA model to estimate the number of human-caused fires in Canada for upcoming years. Note that Tableau uses exponential smoothing to compute the forecasts and present a smoother trend.

The number of human-caused fires in the year of 2023 has been estimated to be 2017 fires around the country.

Recommendation

  • Since human-caused fires are among the two top causes for large-size fires, we need to restrict the barbeque in forests and make a plan to inform people so that we are able to minimize Human-caused fires, especially on days with higher wind gust speed. This will reduce the challenges of forest management for CFS.

  • In most cases, there was a large distance between the fire and the weather stations. In other words, the data might suffer from measurement bias, since there was no closer weather station to the fire. In addition to that, the data has some visible outliers. As observed in the distribution plot, most of the data (“size” variable) lies in the small values.

  • In order to improve the model which have been used in forecasting, in the next steps we need to check and see if we can find data sources for other provinces' weather conditions as well. We also need to find out what the best range of distance is between the fire location and the weather station (after consulting with subject matter experts). This can be applied to the computation and will improve the measurements significantly.


Deliverables