Abstract
The debris flows are one of the most serious geological hazards frequently occurring in the arid and semi-arid regions of northwest China, effectively monitoring and warning against which is of crucial importance. Focusing on the most hazardous debris flow region of Altay, 22 rainfall events that triggered debris flows and 100 rainfall events that did not result in debris flows were systematically evaluated. The mathematical statistics were adopted to identify the characteristics of rainfall (i.e., rainfall intensity, duration, antecedent effectiveness, and direct rainfall amount before the triggering of debris flows). Subsequently, various machine learning methods were employed to determine the optimal impact weights of these rainfall characteristics. The comprehensive rainfall intensity threshold (C(50)) and the difference (C(90)) were derived upon the analysis of the rainfall intensity-duration (I-D) and rainfall intensity-antecedent effective rainfall total (I-E) with the consideration of rainfall characteristics. The accuracy of the warning model was then verified by the comparison with existing monitoring data. It is predicated that there were 36 events with a probability of over 50%, however, only 18 debris flow were actually occurred with a false alarm rate of 50%. For these events with a probability of over 90%, 8 actual debris flow events was occurred within 9 predicated events, the false alarm rate of which is only 11.1%. It is thus believed that that both the accuracy and performance of debris flow prediction can be achieved if the multiple rainfall characteristics are regarded as the auxiliary factors.