Extracting more features from rainfall data to analyze the conditions triggering debris flows

• Published in: Landslides Vol. 19; no. 9; pp. 2091 - 2099
• Main Authors: Zhao, Yan, Meng, Xingmin, Qi, Tianjun, Chen, Guan, Li, Yajun, Yue, Dongxia, Qing, Feng
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2022; Springer Nature B.V
• Subjects: Accuracy; Agriculture; Catchment area; Civil Engineering; Data analysis; Data processing; Debris flow; Detritus; Earth and Environmental Science; Earth Sciences; False alarms; Feature extraction; Geography; Hydrologic data; Landslides & mudslides; Machine learning; Model accuracy; Modelling; Monitoring; Natural Hazards; Original Paper; Precipitation; Precipitation monitoring; Prediction models; Rainfall; Rainfall data; Rainfall intensity; Risk analysis; Risk management; Risk reduction; Statistical analysis; Statistical methods; Early warning; Rainfall features; Debris flows; Triggering conditions
• ISSN: 1612-510X; 1612-5118
• DOI: 10.1007/s10346-022-01893-9

The early warning of debris flows based on rainfall monitoring is an important method for reducing risk. Statistical analysis of the relationship between rainfall characteristics and debris flow events is one of the principal methods of building a threshold model. Therefore, extracting more features from rainfall data is important for improving the prediction accuracy of debris flows. Based on rainfall monitoring records for the Goulinping debris flow catchment in central China, 46 rainfall events triggering debris flows and 321 rainfall events not triggering debris flows were analyzed. One hundred forty-three rainfall features were extracted from the rainfall data using tsfresh (a time series data processing Python package). Using a machine learning method, we tested the accuracy of the model in predicting debris flows in the two cases of using rainfall intensity–duration ( I – D ) (2 features), and using all features (143 features). The results showed that the accuracy of the model using I – D features was 0.94, and the AUC was 0.938; while the accuracy of the model using 143 features was 0.98, and the AUC was 0.975. Calculation of the importance of the features indicated that 7 rainfall features are important in determining the prediction accuracy of the model. The two most important features are the absolute energy (the sum over the squared values of each rainfall value in a rainfall event) and the sum values (the cumulative rainfall). Together they account for 91.4% of the total importance, and they can correctly identify 82.6% of the rainfall events that triggered debris flows without false alarms. Our results indicate that the use of more rainfall features can improve the accuracy and performance of the debris flow prediction model, which has an important reference value for improving the early warning of debris flows.