Abstract

There were 1500 rivers and streams in Taiwan with potential sources of debris flows. Debris flows often occurred after earthquakes and heavy rains, leading to severe damage to human lives and properties. This study aimed to develop an accurate neural network-based risk assessment model of debris flows. The back propagation network was adopted because of its supervised training nature and of its ability to solve complex pattern-matching problems. Real cases of debris flows that occurred in Hualien area of Taiwan from 2007 to 2008 were taken as the database. Such updated data were necessary for accurate predictions of debris flows since the hydrological and geologic data could vary as time proceeds. According to related documentation, this study selected 6 influential factors, including average gradient, catchment area, effective catchment area, accumulated rainfall, rainfall intensity, and geologic condition, as input variables. The results proved that the established model was quite suitable for debris flows risk assessment; the obtained normalized relative error was 8.56% and reduced to 7.04% if geographically close sites were collected as groups.

Key words: Back propagation, debris flow, geographic grouping, neural network, risk assessment.