Spatio-temporal landslide forecasting using process-based and data-driven approaches: A case study from Western Ghats, India

Abstract

The number of rainfall-induced landslides and the resulting casualties are increasing worldwide. Efficient Landslide Early Warning Systems (LEWS) are the best way to reduce the risk due to such events, but the number of operational LEWS is still limited. A new data-driven approach for spatio-temporal landslide forecasting on a regional scale is proposed, integrating Landslide Susceptibility Maps (LSMs) using RF algorithm and probabilistic hydro-meteorological thresholds, considering both rainfall severity and antecedent soil wetness. The proposed method is also compared with two deterministic process-based approaches: Transient Rainfall Infiltration and Grid-based Regional Slope Stability (TRIGRS) and SHALSTAB, considering the spatial variability in soil thickness and properties, along with the rainfall data. The quantitative comparison is carried out for two test areas in the Western Ghats of India (Idukki and Wayanad), for two different spatial resolutions. The efficiency and area under the curve (AUC) values from a receiver operating characteristic curve (ROC) were used to evaluate the performance of different models. The results for Idukki indicate that the efficiency values of the data-driven approach were improved by 4.67 % by using fine resolution DEM (digital elevation model) of 12.5 m resolution, while in the case of TRIGRS and SHALSTAB models, the improvements were 3.39 % and 1.83 %, respectively. For Wayanad, the improvement in efficiencies was further lesser, 2.59 % in the case of data-driven model, and 0.95 % and 0.73 % in the cases of TRIGRS and SHALSTAB, respectively. The maximum efficiency and AUC values were obtained by the data-driven model for both regions, with a spatial resolution of 12.5 m. The maximum efficiency values were obtained as 81.21 % and 83.33 % for Idukki and Wayanad, respectively, and the corresponding AUC values were 0.92 and 0.93. The results indicate that the model proposed in this study, with data-driven approach performs better than the process-based approaches and can bypass the complexities involved in modeling.