A modified Logit model for assessment and validation of debris-flow susceptibility

The accuracy of a debris-flow susceptibility assessment model can not be achieved to an acceptable level without rigorously selecting/processing model input parameters and reliably validating the results with ground truth. In this study, a modified logistic regression model was applied to examine the debris-flow susceptibility in the semi-arid mountainous areas of the southeast Tibetan Plateau, taking the case of the Benzilan-Changbo segment of the Jinsha River. The logistic regression model validation was supported by a field investigation and interpreting aerial photographs. The frequency ratio method was adapted for the sensitivity analysis of ten predisposing factors, including a factor of human activities that has never been considered in previous studies for the selected study area. The model with the highest prediction accuracy (83%) was obtained by comparing 55 possible combinations of the ten factors evaluated using a receiver operating characteristic curve technique. Composite analysis and random sample testing were also conducted to verify the reliability of the susceptibility assessment. The prediction results showed that the actual debris-flow area accounted for 92% of the total predicted debris-flow area. In summary, this study reveals that: (1) the highest susceptibility area was mainly distributed in the northeast and middle riparian zones, partially scattered in the southwest. Human activity was found to be closely related to the probability of debris flow occurrences; (2) under the two situations of independent and index combination participation evaluation, the performances of the same related factors showed largely different effects on the assessment results, which reflected the interrelation and interaction of the factors causing the debris flow; (3) the accuracy of the modified logistic regression model increased by 15%, when compared with the index entropy model in a previous study, and the reasons for this were discussed in the current study. The modified model, along with its enhanced validation process, could be easily extended to other regional mass movement susceptibility analyses.