An analytical solution for rapidly predicting post-fire peak streamflow for small watersheds in southern California

Following wildfires, the probability of flooding and debris flows increase, posing risks to human lives, downstream communities, infrastructure, and ecosystems. In southern California (USA), the Rowe, Countryman, and Storey (RCS) 1949 methodology is an empirical method that is used to rapidly estimate post-fire peak streamflow. We re-evaluated the accuracy of RCS for 33 watersheds under current conditions. Pre-fire peak streamflow prediction performance was low, where the average R-2 was 0.29 and average RMSE was 1.10 cms/km(2) for the 2- and 10-year recurrence interval events, respectively. Post-fire, RCS performance was also low, with an average R-2 of 0.26 and RMSE of 15.77 cms/km(2) for the 2- and 10-year events. We demonstrated that RCS overgeneralizes watershed processes and does not adequately represent the spatial and temporal variability in systems affected by wildfire and extreme weather events and often underpredicted peak streamflow without sediment bulking factors. A novel application of machine learning was used to identify critical watershed characteristics including local physiography, land cover, geology, slope, aspect, rainfall intensity, and soil burn severity, resulting in two random forest models with 45 and five parameters (RF-45 and RF-5, respectively) to predict post-fire peak streamflow. RF-45 and RF-5 performed better than the RCS method; however, they demonstrated the importance and reliance on data availability. The important parameters identified by the machine learning techniques were used to create a three-dimensional polynomial function to calculate post-fire peak streamflow in small catchments in southern California during the first year after fire (R-2 = 0.82; RMSE = 6.59 cms/km(2)) which can be used as an interim tool by post-fire risk assessment teams. We conclude that a significant increase in data collection of high temporal and spatial resolution rainfall intensity, streamflow, and sediment loading in channels will help to guide future model development to quantify post-fire flood risk.

Automated summary

The accuracy of the RCS method was found to be low, underestimating peak streamflow both before and after wildfires. Machine learning techniques were used to develop more accurate models for predicting post-fire peak streamflow, demonstrating the importance of data availability in improving flood risk assessment models.