Enhancing co-seismic landslide susceptibility, building exposure, and risk analysis through machine learning

Ajaya Pyakurel; Diwakar K C; Bhim Kumar Dahal

doi:10.1038/s41598-024-54898-w

Enhancing co-seismic landslide susceptibility, building exposure, and risk analysis through machine learning

Sci Rep. 2024 Mar 11;14(1):5902. doi: 10.1038/s41598-024-54898-w.

Authors

Ajaya Pyakurel¹, Diwakar K C², Bhim Kumar Dahal^{3

4}

Affiliations

¹ Department of Civil Engineering, IOE, Pulchowk Campus, TU, Lalitpur, Nepal.
² Department of Civil and Environmental Engineering, University of Toledo, Toledo, OH, 43606, USA.
³ Department of Civil Engineering, IOE, Pulchowk Campus, TU, Lalitpur, Nepal. bhimd@pcampus.edu.np.
⁴ Institute of Hazard, Risk and Resilience, Durham University, Durham, UK. bhimd@pcampus.edu.np.

Abstract

Landslides are devastating natural disasters that generally occur on fragile slopes. Landslides are influenced by many factors, such as geology, topography, natural drainage, land cover, rainfall and earthquakes, although the underlying mechanism is too complex and very difficult to explain in detail. In this study, the susceptibility mapping of co-seismic landslides is carried out using a machine learning approach, considering six districts covering an area of 12,887 km² in Nepal. Landslide inventory map is prepared by taking 23,164 post seismic landslide data points that occurred after the 7.8 MW 2015 Gorkha earthquake. Twelve causative factors, including distance from the rupture plane, peak ground acceleration and distance from the fault, are considered input parameters. The overall accuracy of the model is 87.2%, the area under the ROC curve is 0.94, the Kappa coefficient is 0.744 and the RMSE value is 0.358, which indicates that the performance of the model is excellent with the causative factors considered. The susceptibility thus developed shows that Sindhupalchowk district has the largest percentage of area under high and very high susceptibility classes, and the most susceptible local unit in Sindhupalchowk is the Barhabise municipality, with 19.98% and 20.34% of its area under high and very high susceptibility classes, respectively. For the analysis of building exposure to co-seismic landslide susceptibility, a building footprint map is developed and overlaid on the co-seismic landslide susceptibility map. The results show that the Sindhupalchowk and Dhading districts have the largest and smallest number of houses exposed to co-seismic landslide susceptibility. Additionally, when conducting a risk analysis based on susceptibility mapping, as well as considering socio-economic and structural vulnerability in Barhabise municipality, revealed that only 106 (1.1%) of the total 9591 households, were found to be at high risk. As this is the first study of co-seismic landslide risk study carried out in Nepal and covers a regional to the municipal level, this can be a reference for future studies in Nepal and other parts of the world and can be helpful in planning development activities for government bodies.

Keywords: Co-seismic landslide; Disaster risk management; Exposure; Machine learning; Susceptibility mapping.