A comparison of statistical and machine learning methods for debris flow susceptibility mapping

• Published in: Stochastic environmental research and risk assessment Vol. 34; no. 11; pp. 1887 - 1907
• Main Authors: Liang, Zhu, Wang, Chang-Ming, Zhang, Zhi-Min, Khan, Kaleem-Ullah-Jan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2020; Springer Nature B.V
• Subjects: Aquatic Pollution; Bayesian analysis; Chemistry and Earth Sciences; Computational Intelligence; Computer Science; Contingency; Datasets; Debris flow; Detritus; Discriminant analysis; Earth and Environmental Science; Earth Sciences; Environment; Flow mapping; Land use; Land use management; Land use planning; Learning algorithms; Machine learning; Mapping; Math. Appl. in Environmental Science; Model accuracy; Mountain regions; Mountainous areas; Original Paper; Performance evaluation; Physics; Probability Theory and Stochastic Processes; Regression analysis; Risk reduction; Statistical analysis; Statistics for Engineering; Susceptibility; Waste Water Technology; Water Management; Water Pollution Control; Himalayas area; Logistic regression; Random forest; Debris flow; Bayes discriminant analysis; Susceptibility
• ISSN: 1436-3240; 1436-3259
• DOI: 10.1007/s00477-020-01851-8

Debris flows destroys the facilities and seriously threatens human lives, especially in mountainous area. Susceptibility mapping is the key for hazard prevention. The aim of the present study is to compare the performance of three methods including Bayes discriminant analysis (BDA), logistic regression (LR) and random forest (RF) for debris flow susceptibility mapping from three aspects: applicability, analyticity and accuracy. Nyalam county, a debris flow-prone area, located in Southern Tibet, was selected as the study area. Firstly, the dataset containing 49 debris flow inventories and 16 conditioning factors was prepared. Subsequently, divided the dataset into two groups with a ratio of 70/30 for training and validation purposes, and repeated 5 times to obtain 5 different groups. Then, 16 factors were involved in the modeling of RF, of which 11 factors with low linear correlation were for BDA and LR. Finally, receiver operating characteristic curves, the area under curve (AUC) and contingency tables were applied to evaluated the accuracy performance of the 3 models. The prediction rates were 74.6–81.8%, 74.6–83.6% and 80–92.7%, for the BDA, LR and FR, while the AUC values of three models were 0.72–0.78, 0.82–0.92 and 0.90–0.99, respectively. Compare to LR an BDA, RF not only effectively process and preserved dataset without priori assumption and the obtained susceptibility zoning map and major factors were reasonable. The conclusion of the current study is useful for risk mitigation and land use planning in the study area and provide related references to other researches.