Rainfall threshold for landslide activity in Dazhou, southwest China

• Published in: Landslides Vol. 17; no. 1; pp. 61 - 77
• Main Authors: Yang, Hongjuan, Wei, Fangqiang, Ma, Zhenfeng, Guo, Haiyan, Su, Pengcheng, Zhang, Shaojie
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 2020; Springer Nature B.V
• Subjects: Agriculture; Civil Engineering; Daily precipitation; Datasets; Earth and Environmental Science; Earth Sciences; Empirical analysis; Geography; Landslide effects; Landslide warnings; Landslides; Natural Hazards; Original Paper; Probability theory; Rain; Rainfall; Regression analysis; Slope stability; Statistical analysis; Thresholds; Warning; China; Logistic regression; Landslide; ROC analysis; Rainfall threshold
• ISSN: 1612-510X; 1612-5118
• DOI: 10.1007/s10346-019-01270-z

Dazhou is one of the landslide-prone areas in southwest China. Since rainfall is the primary factor responsible for slope instability in this region, empirical rainfall thresholds may be effective for landslide warning. Thresholds using the intraday rainfall and the effective antecedent rainfall up to 7 days prior to landslide occurrence have been defined for this area. The study region was divided into 6 alert zones. For each zone, the daily rainfall and landslide data in the monsoon season during the period of 2003–2010 were analyzed, and the logistic regression method was employed to fit the dataset and provide diverse probability thresholds. Subsequently, the optimal threshold was identified by receiver operating characteristic (ROC) analysis. Furthermore, two classes of criteria and the associated indicators were used to determine the three rainfall thresholds for the warning model with four classes of alert (i.e., no warning, moderate warning, high warning, and very high warning). Finally, these thresholds were validated with dataset for the period 2011–2015. It suggests that the optimal threshold derived from ROC analysis, with the probability ranging from 4 to 10%, is suitable for issuing moderate warning, while the probability corresponding to the thresholds for high warning and very high warning varies in the ranges 30–55% and 75–95%, respectively. Performance of these thresholds in the period 2011–2015 is comparable to that in the period 2003–2010, indicating the robustness of them and the potential for practical use.