Debris Flow Prediction Model Based On Bayes Discriminant Analysis In Sichuan Province

In the research of debris flow, the prediction is still an important research direction.Although theoretical methods of prediction are developing continually, themathematical statistics methods are still importance. Bayes discriminant analysismethod, rarely seen to be used in the domestic debris flow prediction, is worth of furtherstudy. Debris flow involves many related disciplines, where hydrology and environmentare more prominent. The quick development of remote sensing (RS) and geographicinformation system (GIS) technology provides a reliable support for debris flowprediction study.In the study, first, spatial interpolation optimization considering environmentalfactors is to estimate better quality daily rainfall in Sichuan Province; secondly, Bayesdiscriminant analysis models were established based on the different combinations ofthe rainfall and environment factors in different areas; finally, under different priorprobability combinations of debris flows, two types of prediction models were built inthe light of Bayes discriminant analysis (BDA) and Logistic regression (LR), and then acomparison was made between their prediction accuracy which was got by trainingsample. The main results are as follows.(1) For rainfall spatial interpolation optimization, in co-Kriging interpolationprocess, elevation, slope, and aspect were introduced as auxiliary information, whichwere selected by means of correlation analysis and principal component analysis. Andthen compared with Ordinary Kriging by means of crosscheck, the co-Kriging wasslightly better on the whole.(2) According to relationship analysis of debris flow in Sichuan, the appropriaterainfall and environment were selected as predictors of models, and to build the modelsof BDA in the whole province area and the high incidence area, respectively. Thediscriminant accuracy rate was obtained through itself validation and cross-validation.The results were that (a) with rainfall or rainfall and environment as predictors, theoverall discriminant accuracy rate of high incidence areas were2.75%higher and5.9%higher than those of the whole areas on average, respectively;(b) in the entire areas, the prediction model based on the rainfall and environment factors with0.9%difference ofthe overall discriminant accuracy rate between itself validation and cross-validation,was more stable than that based on the rainfall factors with2.8%difference, while in thehigh incidence area, the prediction model based on the rainfall and environment factorswith2.7%difference was less stable than that based on the rainfall factors with1.1%difference.(3) In consideration of debris flow priori probability, on the basis of the above, Panzhihua and Liangshan Yi Autonomous Prefecture were selected as the study areas tomodel and compare the BDA and LR prediction, giving priority to with the overallpredicting accuracy rate in sample. The results showed that (a) with rainfall andenvironment factors been used as the predictors, when the priori probabilitycombinations of occurrence and nonoccurrence were in order of (0.5,0.5),(0.67,0.33),(0.75,0.25) and (0.9,0.1), the overall predicting accuracy rate of BDA were8.3%higher,3.2%higher,0%higher and0.6%lower than those of LR;(b) with rainfallfactors as the predictors, when the priori probability combinations of occurrence andnonoccurrence were in order of (0.5,0.5),(0.67,0.33),(0.75,0.25) and (0.9,0.1), theoverall predicting accuracy rate of BDA were0.5%lower,4.7%higher,0.5%higherand0.5%lower than those of LR.Synthetically considering the above results, building the Bayes discriminantpredicting models of the different condition combination was an effective and simpleway on the grounds of the actual debris flow occurrence in Pan zhihua and Liangshanregion, which had some practical value.