Human driven climate change increased the likelihood of the 2023 record area burned in Canada

Abstract

In 2023, wildfires burned 15 million hectares in Canada, more than doubling the previous record. These wildfires caused a record number of evacuations, unprecedented air quality impacts across Canada and the northeastern United States, and substantial strain on fire management resources. Using climate models, we show that human-induced climate change significantly increased the likelihood of area burned at least as large as in 2023 across most of Canada, with more than two-fold increases in the east and southwest. The long fire season was more than five times as likely and the large areas across Canada experiencing synchronous extreme fire weather were also much more likely due to human influence on the climate. Simulated emissions from the 2023 wildfire season were eight times their 1985-2022 mean. With continued warming, the likelihood of extreme fire seasons is projected to increase further in the future, driving additional impacts on health, society, and ecosystems.

Keywords: Attribution; Climate change; Climate-change impacts.

Fig. 1. 2023 area burned in Canada.

a Map of Canada with area burned in 2023 shown in red. Forested area is shaded in light green. Ecozone boundaries are in bold and province/territory boundaries in gray. b Ecozone names, abbreviations, 2023 total area burned in each region, expressed in thousands of hectares (kha), and rank in the time series 1972–2023. c The long-term record of national area burned, expressed in millions of hectares (Mha).

Fig. 2. Regression models were tested and fit using ERA5 reanalysis data and observed area burned and then the best model was applied to the climate model ensembles.

a For each ecozone (column) and potential predictor (row) from ERA5, the variance explained (R²) value is shown from a regression fit with the log of the observed area burned for 1972-2022. Stars indicate the predictor with the largest R² for each region. For each index, additional details are listed in the table to the left, including the name and abbreviation, daily input (T: temperature, RH: relative humidity, ws: wind speed, pr: precipitation), and metric used to summarize to annual values. For the metrics, p95 refers to the 95^th percentile, JJA the mean (or sum in the case of cDSR) across the fire season, and 7X refers to the fire-season maximum of the seven-day running means (15X and 31X refer to the maximum of the 15-day and 31-day running means, respectively). An extended version of (a) is available in Supplementary Fig. 1 including more predictors. Time series (b) of regression-predicted area burned (log transform) for the CMIP6-historical ensemble for each ecozone. The CMIP6-historical ensemble mean is in bold blue and the shading indicates the uncertainty range, considering both model spread and regression error of prediction. The observed area burned is shown in black with a dotted line at the 2023 value.

Fig. 3. Attribution of the 2023 area burned and fire weather.

Change in likelihood of an event at least as large as the 2023 (a) area burned, b maximum seven-day-mean FWI, and (c) fire season length by ecozone that is attributable to human influence on the climate. Each region (see Fig. 1 for the long names) is shaded according to the lower bound of the risk ratio (the ratio of the probability of the 2023 value with all forcing to the probability of the observed event in the counterfactual), based on results using CMIP6-historical. The 2023 value was extreme in many, but not all, regions. Below each region label is a set of three dots, shaded according to the same color scale, to indicate consistency of the results. These dots indicate the risk ratio using, left to right: CMIP6-DAMIP, CanLEAD-FWI, CMIP6-HighResMIP. See Supplementary Fig. 2 for the best estimate of the risk ratio.

Fig. 4. Attribution of the cumulative high fire risk area.

a Time series of the cumulative high fire risk area calculated as the sum across the fire season of the daily forested area experiencing a Fire Weather Index value greater than its local 95th percentile based on a 1951–1980 climatology. ERA5 is shown in black and CMIP6-historical forcing in green with the ensemble means in bold and the 5th to 95th percentile range shaded. b Risk ratios (RR) for a cumulative area at least as large as observed in 2023 for each ensemble. The triangle for CMIP6-HighResMIP indicates an infinite RR and the upper bounds on the uncertainty range extend to infinity for CMIP6-historical and CanLEAD-FWI. An infinite RR occurs when the event in question was very rare and did not occur in the counterfactual simulations (or the resampling of them) and can be interpreted to mean that all of the likelihood of the event’s occurrence is due to human influence on the climate.

Fig. 5. Land surface model (LSM) based fire emissions estimates.

Historical fire emissions from the CLASSIC LSM from 1985 to 2023. The dashed line denotes the 1985 to 2022 LSM historical average tied to the right y-axis. The shaded region represents the standard deviation from the mean for two LSM runs using two driving meteorologies. The red-shaded region represents the annual range of five independent historical gridded estimates. ^aData sets include the Global Fire Emissions Database version 4.1 with small fires, the Fire Inventory from NCAR version 2.5, Fire Energetics and Emissions Research version 1.0-G1.2, the Quick Fire Emissions Dataset version 2.4 revision 1, and Carbon Tracker 2019.

Fig. 6. Change in likelihood of area burned for a future climate.

Similar to Fig. 3a, the lower bound on the uncertainty range of the risk ratio for an area burned at least as large as that observed in 2023 is shown but for a climate with a global temperature +3 °C warmer than pre-industrial. The results shown here are using the CanLEAD-FWI dataset and compare to the center circle in Fig. 3a for the +1 °C global warming climate.