VERIFICATION AND TESTING OF MODELS FOR POST-WILDFIRE DEBRIS FLOW PROBABILITY

Frequent and extensive wildfires in southern California effectively alter the soil properties and surface hydrology on steep slopes, thereby increasing the probability of debris-flows and the resulting threat to life and property. Twenty-seven previously derived, probability models for post-fire, debris-flow occurrence were assessed using the Chung and Fabbri method (2008) to find the most representative for the southern California study area. This method calculates two curves for each model; Success Rate Curves are generated using the database used to develop the models, and Prediction Rate Curves are calculated using an independent dataset. All but one of the 27 models resulted in Success Rate Curves that showed 90% of the debris flows occurring in basins where at least a 60% probability was calculated. This finding, combined with the statistical measures evaluated during model-building, indicates that 26 of the 27 models can provide a reasonable estimate of debris flow probability. However, some models showed better distribution of values between 1 and 100% probability than did others. The predictive capabilities of the models were tested by producing Prediction Rate Curves from 86 known responses for the 2008-2009 storm season. Because the dataset for the recent season included only 7 observed debris-flow events and 3 storms, its utility in testing the Prediction Rate Curves was limited.

Consideration of the results from both the Success and Prediction Rate Curves indicated that the most reasonable probability model includes the length of the longest flow path within the basin, total basin relief, the percentage of the basin burned at high and moderate severity, the percentage of the burned basin burned with slopes greater than or equal to 30%, the percentage of clay in the soil, the total storm duration, and the average storm intensity. Although this model will currently be helpful in assessing hazards the predictive strength of the probability models could be improved by separating the model database and producing at least two separate models; one for long-duration, low-intensity frontal storms and on for short-duration, high-intensity frontal storm responses.

References: Chung & Fabbri, Predicting landsides for risk analysis – Spatial models tested by a cross-validation technique, Geomorphology 94, 2008, (438-452)