POST-FIRE DEBRIS FLOW PREDICTION USING A TWO-STEP HYBRID APPROACH

The prediction of debris flow generation from burned basins is not easily conducted using traditional modeling approaches. This study explores the efficacy of using an alternative modeling paradigm based on a two-step hybrid approach. First, data mining by the self organizing map technique is used to identify statistically significant variables comprising conceptual multistate models of the western United States. Second, the conceptual model variables are used as information from which the genetic programming technique discovers predictive post-fire debris-flow peak discharge and total volume equations based on evolutionary heuristics. The search space is constrained using a multi-component objective function that simultaneously maximizes fitness and minimizes root-mean-squared and unit errors for the discovery of fittest equations. Equations associated with the lowest root-mean-squared-error (RMSE) values tend to be physically unrealistic and associated with the largest unit errors. By accepting larger RMSE values, the equations are physically realistic, parsimonious with respect to the function set, and dimensionally correct. In contrast to the published multiple linear regression (MLR) equations, the hybrid modeling approach discovers equations whose predictions of post-fire debris flow peak discharge and total volume are unbiased and better related to observations, and have less prediction uncertainty. For example, the estimated minimum to maximum total volume prediction uncertainty for the MLR equation spans a factor of about 6, whereas the average for discovered equations spans a factor of about 2. Further reductions in prediction uncertainty may be possible when dimensional consistency is not a priority and by subsequently applying a gradient solver to fittest solutions.