Abstract. Rainfall-induced debris flows in recently burned mountainous areas cause
significant economic losses and human casualties. Currently, prediction of
post-fire debris flows is widely based on the use of power-law thresholds and
logistic regression models. While these procedures have served with certain
success in existing operational warning systems, in this study we investigate
the potential to improve the efficiency of current predictive models with
machine-learning approaches. Specifically, the performance of a predictive
model based on the random forest algorithm is compared with current techniques
for the prediction of post-fire debris flow occurrence in the western United
States. The analysis is based on a database of post-fire debris flows
recently published by the United States Geological Survey. Results show that
predictive models based on random forest exhibit systematic and considerably
improved performance with respect to the other models examined. In addition,
the random-forest-based models demonstrated improvement in performance with
increasing training sample size, indicating a clear advantage regarding their
ability to successfully assimilate new information. Complexity, in terms of
variables required for developing the predictive models, is deemed important but
the choice of model used is shown to have a greater impact on the overall
performance.
Updated logistic regression equations for the calculation of post-fire debris-flow likelihood in the western United States