Evaluation and validation of multiple predictive models applied to post-wildfire debris-flow hazards.

摘要

The combination of greater wildfire frequency and expansion of populations into the wildland-urban interface drives a need for accurate prediction of landslides and debris-flow hazards. Statistical methods are frequently used to rapidly assess landslide and debris-flow hazards for emergency planning and risk assessment. The U.S. Geological Survey is working to improve empirical logistic regression models for post-wildfire debris-flow probability and expand the regions in which they can be applied. Numerous methods have been used to evaluate predictive logistic regression models and there is no consistent approach existing in the landslide or debris-flow literature. There is a need to develop recommendations for the evaluation and comparison of multiple predictive models. This research attempts to address this need by evaluating the predictive performance of debris-flow likelihood models using a combination of statistical and objective measures. Regression evaluation statistics were used to identify the top-performing predictive models on a training dataset. Classification performance metrics were used to compare predictive success between different models on both a training and test dataset. In addition, model performance was evaluated on a regional scale based on the Koppen climate classification. Although this research focused on post-wildfire debris-flow prediction, the methodology is applicable to any probability-based binary classifier model, and can be used to evaluate predictive models that address a wide range of natural hazards. The systematic framework established in this research uses statistical and objective measures to guide the selection, evaluation, comparison, and validation of multiple binary predictive models. The recommendations from this work should provide a consistent approach and identify necessary reporting elements for presenting prediction models in geologic literature.