Temporal Trends in Forest Fire Season Length Alisha Albert-Green aalbertg@sfu.ca Department of Statistics and Actuarial Science Simon Fraser University Stochastic Modelling of Forest Dynamics Webinar March 9, 2011 1 / 24
Outline Introduction Methodology Alberta Results Ontario Results Conclusions 2 / 24
Introduction On average, each in Canada: 8000 fires 2.1 million ha burned $500 million - $1 billion spent in suppression costs (Natural Resources Canada) Annually, lightning-caused fires account for approximately: 45% of ignitions 80% of area burned (Natural Resources Canada) 3 / 24
Introduction Increasing temperatures could: increase number of ignitions increase amount of severe fire-weather extend fire season (Weber & Stocks, 1998) Woolford et al. (2010): used logistic generalized additive mixed models to investigate possible climate change effects in lighting-caused fires for a region in Northwestern Ontario observed seasonal and annual changes in ignition risk 4 / 24
Introduction Goals: Data: Predict fire day risk Estimate the length of the fire season Test for significant temporal trends in the length of the fire season Lightning-caused fires from: Alberta, Canada (1961-2003) Ontario, Canada (1963-2004) 5 / 24
Statistical Models Linear Regression Models: E[y x] = β 0 + β 1 x Generalized Linear Models: g(µ) = β 0 + β 1 x where µ = E[y x], y exponential family distribution Generalized Additive Models (GAMs): incorporate non-linear relationships via smooth functions g(µ) = β 0 + f (x) 6 / 24
Logistic GAM Framework 1: bivariate smooth model logit[p(t dy )] = f 1 (d, y) where: p(t dy ) = risk of a fire day ( ) p logit(p) = log 1 p d = day of y = 7 / 24
0.4 Estimated Risk of a Fire Day for Alberta: bivariate smooth model estimated risk 0.6 0.4 0.2 2000 1990 1980 1970 100 150 200 300 250 day of 1970 1980 1990 2000 0.1 0.3 0.2 0.5 0.7 0.6 0.5 0.4 0.2 0.3 0.1 100 150 200 250 300 day of (a) Surface plot. (b) Contour plot. 8 / 24
Logistic GAM Framework 2: univariate smooth model logit[p(t dy )] = f 2 (t) t = {1 = January 1 1961, 2 = January 2 1961,..., T = December 31 2003} Estimated Risk for Alberta Fires estimated risk 0.8 0.6 0.4 0.2 0.0 9 / 24
Estimating the Length of the Fire Season: example from 1965 with a 0.05 threshold 0.6 0.5 estimated risk 0.4 0.3 0.2 0.1 0.0 Jan 1 Mar 1 May 1 July 1 Sep 1 Nov 1 10 / 24
Estimating the Length of the Fire Season: example from 1965 with a 0.05 threshold 0.6 0.5 estimated risk 0.4 0.3 0.2 0.1 0.0 Jan 1 Mar 1 May 1 July 1 Sep 1 Nov 1 11 / 24
Estimating the Length of the Fire Season: example from 1965 with a 0.05 threshold 0.6 0.5 estimated risk 0.4 0.3 0.2 0.1 0.0 Jan 1 Mar 1 May 1 July 1 Sep 1 Nov 1 12 / 24
Estimating the Start of the Fire Season: example with a 0.05 threshold goal: quantify trend over the s May 1 Apr 15 Apr 1 13 / 24
Estimating the Start of the Fire Season: example with a 0.05 threshold goal: quantify trend over the s May 1 Apr 15 Apr 1 14 / 24
Estimating the Start of the Fire Season: example with a 0.05 threshold goal: quantify trend over the s May 1 Apr 15 Apr 1 15 / 24
Weighted Regression Models: example using a 0.05 threshold May 1 Apr 15 Apr 1 16 / 24
Weighted Regression Models: beginning of the fire season May 1 Apr 15 Apr 1 Mar 15 Mar 1 Feb 15 Feb 1 Jan 15 Jan 1 Jun 1 May 1 Apr 15 (a) Threshold of 0.01. (b) Threshold of 0.10. 17 / 24
Weighted Regression Models: end of the fire season Nov 1 Dec 15 Dec 1 Nov 15 Nov 1 Oct 15 Oct 1 Sep 15 Sep 1 Oct 15 Oct 1 Sep 15 Sep 1 Aug 15 (a) Threshold of 0.01. (b) Threshold of 0.10. 18 / 24
Empirical Analysis of Alberta Fires: beginning of the fire season Fifth Observed Fire Day Per Year First Day the Empirical CDF Exceeds 5 % Jun 1 May 1 Jun 15 Jun 1 First Observed Day Per Year There Are At Least 2 Fires First Day There Are Two Fire Days Within Four Days Jun 15 Jun 1 May 1 Jun 1 May 1 19 / 24
Empirical Analysis of Alberta Fires: end of the fire season Oct 15 Oct 1 Sep 15 Sep 1 Aug 15 Aug 1 Fifth Last Observed Fire Day Per Year Sep 15 Sep 1 Aug 15 Aug 1 First Day the Empirical CDF Exceeds 95 % Last Observed Day Per Year There Are At Least 2 Fires Last Day There Are Two Fire Days Within Four Days Sep 15 Sep 1 Aug 15 Aug 1 July 15 Oct 15 Oct 1 Sep 15 Sep 1 Aug 15 Aug 1 20 / 24
Empirical Analysis of Ontario Fires Fifth Observed Fire Day Per Year Fifth Last Observed Fire Day Per Year Jun 15 Jun 1 May 1 Oct 1 Sep 15 Sep 1 Aug 15 1970 1980 1990 2000 1970 1980 1990 2000 July 1 Jun 15 Jun 1 May 1 First Day the Empirical CDF Exceeds 5 % 1970 1980 1990 2000 Oct 1 Sep 15 Sep 1 Aug 15 Aug 1 First Day the Empirical CDF Exceeds 95 % 1970 1980 1990 2000 (a) Beginning of the fire season. (b) End of the fire season. 21 / 24
Weighted Regression Models: 0.05 threshold Jun 1 May 1 Apr 15 Apr 1 Nov 1 Oct 15 Oct 1 Sep 15 Sep 1 1970 1980 1990 2000 1970 1980 1990 2000 (a) Beginning of the fire season. (b) End of the fire season. 22 / 24
Conclusions and Future Work Conclusions Significant trends were found in the length of the fire season in Alberta No significant trends were found in Ontario Future Work Investigating the use of additional covariates: teleconnections such as: Pacific Decadal Oscillation (PDO), El-Niño Southern Oscillation (ENSO), Arctic Oscillation (AO), etc. fire weather covariates such as: fire weather index (FWI), fine fuel moisture code (FFMC), duff moisture code (DMC), etc. Incorporating correlation: temporal correlation spatial correlation 23 / 24
Acknowledgements Financial support from GEOIDE and NSERC is gratefully asknowledged. We thank Alberta Sustainable Resource Development and Ontario Ministry of Natural Resources for the use of their data. 24 / 24