4. Results
4.1 Pre-fire and Post-fire Comparisons
Multi-factor ANOVA was performed on DOC and SUVA in fire-disturbed and fire-undisturbed catchments pre and post the wildfire. The results are displayed in Table 2. Year and fire disturbance were both found to have statistically significant effects on DOC (p-value < 0.001). The sum of squares of year effects is more than twice of the fire disturbance. This indicates the total variation due to year effects is higher than fire disturbance. Year was found to have a statistically significant effect on SUVA (p-value < 0.001). According to the ANOVA table, fire disturbance did not significantly affect the changes of SUVA and its sum of squares appears to be very small.
​
Figure 5 shows the bar plots for DOC and SUVA. In general, DOC and SUVA increased in fire-disturbed catchments compared to fire-undisturbed catchments. Referring to the DOC baseline value (7.9 L/mg), there were noticeable differences of DOC in both fire-disturbed and fire-undisturbed catchments after the fire. Average DOC concentration in fire-undisturbed catchments post fire was 5.8 times greater than the baseline and average DOC in fire-disturbed catchments post fire was 7.4 times greater than the baseline.
Table 2. Summary of multi-factor ANOVA on DOC and SUVA change before and after wildfire.
Figure 5. Bar plots of DOC (mg/L) and SUVA (L/mg·m) in fire-disturbed and fire-undisturbed catchments in year 2020 and 2021 (error bars represent +/- 1 SE).
4.2 Post-fire Patterns
Post-fire patterns of DOC and SUVA in fire-disturbed catchments are observed in Figure 6. Looking at the median, DOC showed a steady increase and then plateaued from May to September 2021. The values became more dispersed and higher concentrations were present in August and September. SUVA experienced the peak in August and its value decreased in September.
Figure 6. Box plots of DOC (mg/L) and SUVA (L/mg·m) in fire-disturbed catchments after the wildfire.
4.3 Feature Importance on Post-fire DOC
A random forest model was implemented on post-fire DOC change in the fire-disturbed catchments. As an initial check, I plotted the correlation matrix of the fifteen predictor variables for evaluating the DOC concentrations. As seen in Figure 7, the correlation coefficients are low between most variables. The proportion of organic deposits and disturbed peat appear positively correlated, and I removed organic deposits for the random forest modeling.
Figure 7. A correlation heatmap showing the relationship between the predictor variables. Correlation coefficients are labelled and a darker color indicates a stronger correlation.
Figure 8 shows the result of the random forest. 77% of the DOC variance was explained by the model, and the RSME was 17.60. Feature importance of the fourteen predictors variables for DOC. The top three important variables were percent of wetland, settlement, and cropland. Percent of fire burned was ranked as the fourth most important.
Figure 8. Random forest model of predicted vs actual DOC (mg/L) and feature importance of land use, surficial geology, and wildfire disturbance used in the model. The higher the score of % increased error, the more important the variable is.
I further calculated the covariance among the top four important predictor variables classified by the random forests. Figure 9 shows their variance partitioning. In total, the four variables explained 53% of the DOC change after the fire and the shared fractions between variables (the overlap) were 24%.
.png)
Figure 9. Venn diagram of the relationship between percent of wetland, settlement, cropland and area burned. The overlap between colors represents covariance. Values smaller than 0 are not shown.
5. Discussion
From the analysis, the DOC in both fire-disturbed and fire-undisturbed catchments showed an increase after the Tomahawk fire, with the magnitude of increases larger in fire-disturbed catchments. The DOC increases in fire-undisturbed catchments were not anticipated and this could be due to the neglect of hydrological connectivity between catchments in this study. Due to the year effects, SUVA values from 2021 was evidently higher than SUVA values from 2020; however, SUVA was not significantly affected by fire disturbance (Table 2). The consistent difference of SUVA in all catchments between year 2020 and 2021 could have suggested large annual variation. Overall, results showed that average post-fire DOC concentration in all fifty-five catchments increased more than five times of the DOC baseline for the Tomahawk region. DOC removal in drinking water treatment involves chemical cost of coagulants, sludge removal, energy cost and treatment plant operational costs (Xu et al., 2020). Larger capital investments would be required in the downstream community to process source water and maintain drinking water safety.
In fire-disturbed catchments, the post-fire DOC showed an increasing trend as expected. On the other hand, post-fire SUVA reached the maximum and then decreased; however, the values were dominantly above 4 L/mg·m. Typically, SUVA greater than 4 L/mg·m is accompanied by high potential of disinfection by-product formation and indicates higher coagulant dosage demand or additional color removal in water treatment (Health Canada, 2020). Increases of coagulant demand imply increased treatment costs. Given the measurements were taken within five months, these insights are valid for short-term water treatment and might not be reflective for long-term consideration.
According to the random forest regression model, three land use variables were ranked more important than the area burned in terms of post-fire DOC change. Davidson et al. (2019) reported similar findings that the hydrogeologic factors could override the influence of wildfire. Looking at the variance partitioning (Figure 9), the variable of fire burned does contribute to explaining the DOC change; meanwhile, it shares 14% variance with wetland, settlement, and cropland variables. This covariance indicates dependency between these four variables and it suggests a mixing influence of wildfire and different landscape variables on DOC change. Given the warming climate, the frequency of wildfires is likely to increase in the future. Fire disturbances could complicate the landscape and impact DOC dynamics more profoundly. It is necessary to take into account different land usage and make management adaptations accordingly to facilitate the protection of drinking water supply. For instance, wetland conservation may be a priority to ensure carbon sequestration (Were et al., 2019). Human interventions could be taken in reducing agricultural runoffs to the nearby catchments.