Research On Related Issues Of Single Ditch Debris Flow Risk Assessment Based On Machine Learning Theory

The risk assessment of debris flow is not only the main content of the pre-disaster assessment of debris flow,but also an important reference for the prevention and control of debris flow.A reasonable assessment of the current state of the debris flow and effective prediction of the future development trend of the debris flow have important practical significance for the monitoring and early warning of debris flow disasters and the formulation of risk avoidance plans.With the development and improvement of the theory of various disciplines,the field of debris flow risk assessment has gradually realized the cross application of various new technologies and methods.Although the results are gratifying on the whole,most scholars pay more attention to the prediction results.Therefore,it is necessary to carry out further research on the key issues such as the characteristics and applicability of the single gully debris flow risk assessment model and evaluation index system determination method,the influence mechanism of the imbalance and spatial variability of debris flow data samples on the prediction accuracy and generalization ability of the model.Therefore,the research on the risk assessment of single gully debris flow based on machine learning in this paper not only helps to realize the cross integration of various disciplines,but also has guiding significance to the selection of evaluation model,the selection of sample data and the construction of index system.Taking Beichuan County,Longchi area of Dujiangyan,Yunnan Province and Jishixia reservoir area of the Yellow River as research samples,based on the theories of artificial neural network,support vector machine,grey correlation degree,decision tree,rough set,principal component analysis,resampling,under sampling and Ada Boost algorithm,this paper systematically and comprehensively analyzes the characteristics and applicability of machine learning prediction models,the characteristics and applicability of evaluation index system construction methods,how to deal with the imbalance of debris flow sample data and the spatial variability of debris flow systems The impact mechanism on the generalization ability of forecasting model,The following research results are obtained:(1)Four key problems in the field of risk assessment of single gully debris flow based on machine learning theory are summarized.At the same time,a reasonable research scheme is formulated for specific problems,and the main research content of this paper is determined.(2)The grey correlation model can be used as the preferred prediction model for single gully debris flow risk assessment;neural network model and support vector machine model have their own advantages,and the accuracy of neural network model prediction results is greatly affected by the degree of sample imbalance,but less affected by the number of samples,support vector machine model is just the opposite;decision tree model in the risk of single gully debris flow The accuracy of degree evaluation is low and it is difficult to be widely used.(3)Rough set theory can be used as the first choice of evaluation index construction method.When the correlation degree of evaluation indexes is low,the evaluation index system determined by grey correlation degree reflects more comprehensive information and the modeling effect is better.When the correlation between evaluation indexes is high,but the evaluation indexes are closely related to the evaluation results,the principal component analysis has more advantages.(4)Resampling method can be used as the first choice to deal with the problem of unbalanced sample data;under sampling method uses a small number of test samples,and the accuracy of the prediction model is affected to a certain extent;when the number of test samples is small,the error sum of BP Ada Boost algorithm is often 1 or 0,resulting in the model can't run normally;if the number of test samples is small,the prediction model can't run normally Because of the large amount of data,the combination of resampling and BP Ada Boost algorithm may have a better application prospect.(5)Spatial variability mainly affects the generalization ability of the model by controlling the sensitivity of the same evaluation index in different regions.The new evaluation index system constructed by principal component analysis can effectively solve the change of evaluation index sensitivity caused by the change of spatial position.