Geological Hazards Susceptibility Evaluation Based on GA‐BPNN: A Case Study of Xingye County

• Published in: Earth and space science (Hoboken, N.J.) Vol. 9; no. 12
• Main Authors: Xu, Kai, Ma, Xiaogang, Liu, Gang, Wu, Chonglong, Kong, Chunfang
• Format: Journal Article
• Language: English
• Published: Hoboken John Wiley & Sons, Inc 01.12.2022; American Geophysical Union (AGU)
• Subjects: back‐propagation neural networks; Fuzzy logic; genetic algorithm; Geographic information systems; Geographical Information Systems; Geography; Geological hazards; geological hazards (geohazards); Geology; Land use; Lithology; Machine learning; Neural networks; Remote sensing; Support vector machines; susceptibility evaluation; Sustainable development; China
• ISSN: 2333-5084; 2333-5084
• DOI: 10.1029/2019EA000929

Geohazards are one of the most critical disasters that cause serious damage to engineering construction activities and peoples' lives and property; therefore, susceptibility evaluation of the geohazards is vital for planning future sustainable development and construction activities. This paper presents a multidisciplinary approach to divide the susceptibility grades of geohazards in Xingye County, using back‐propagation neural networks (BPNN), Geographical Information Systems, and grid analysis technology on remote sensing data. Twelve geohazards related causal factors were selected to reflect the physical geography, basic geology, ecological environment, human activities, and other characteristics of Xingye; the factors are: slope, aspect, topographic curvature, precipitation, normalized differential vegetation index, stratum lithology, linear faults, tectonic complexity, normalized differential water index, residential density, road network density, and land use/land cover. Two hidden layers of BPNN optimized by genetic algorithm (GA) were trained to evaluate the susceptibility grade of geohazards for Xingye. Receiver operating characteristics (ROC) curves, mean square error (MSE), mean of absolute percentage error (MAPE), and field investigation were used to test the validity of the models. Analysis and comparison of the results of different models as indicated by the values of area under curve (AUC) of ROC curves, MSE, and MAPE for five models clearly indicated that the GA‐BPNN model developed in the present study has the highest accuracy (0.9344) and the least errors (0.1925% and 7.3833%) for the susceptibility evaluation and prediction of geohazards in Xingye, followed by the BPNN, support vector machine, random forest, and logistic regression. Moreover, the results also show that the GA has a positive effect on the BPNN model. Furthermore, the GA‐BPNN model developed in the present study can be used for the assessment of geohazards susceptibility in similar regions. Key Points Establishing the evaluation index system for geohazards based on natural ecological environment characteristics and human activities Developing a geohazard susceptibility assessment model based on two hidden layers of back‐propagation neural networks (BPNN) optimized by genetic algorithm (GA) for dividing the geohazard levels Verifying the effectiveness of the geohazard susceptibility assessment model based on GA‐BPNN by analyzing and comparing the values of area under curve, mean square error, and mean of absolute percentage error of the different models