Comparison of the advantages and disadvantages of different machine learning methods in geohazard risk assessment: Taking Da'an Town, Ningqiang County as an example

This paper is the result of geohazard survey engineering.

Objective The occurrence of geohazards are influenced by various factors, which have uncertainty and complexity, making it difficult to assess the risk of geohazards. With the development of AI technology, intelligent algorithms can more accurately calculate the complex and nonlinear relationships between geohazard triggering indexes, greatly improving the accuracy of geological hazard risk assessment models.

Methods Based on the field geological survey data of Da'an Town, Ningqiang County, 12 indexes closely related to the occurrence of geohazards were selected, namely elevation, slope, slope height, slope direction, slope type, engineering geological rock formations, fault distance, water system distance, road distance, vegetation coverage, rainfall, and seismic ground motion, as risk zoning evaluation factors. By constructing a sample set, Bayesian, strategy gradient neural network, random forest, KNN and neural network algorithm are used to model and compare the geohazard risk assessment result in Da'an Town, Ningqiang County.

Results The experimental results show that the Bayesian model (AUC 0.894) performs the best, with the vast majority of geohazards located in the extremely high and high−risk evaluated areas, and meets the requirements for prediction accuracy evaluation.

Conclusions It is feasible to choose Bayesian algorithm models for geological hazard risk assessment when the number of geohazard samples is small.