Landslide susceptibility assessment using maximum entropy model with two different data sampling methods

- MaxEnt has an acceptable performance in landslide susceptibility modelling.
- HAND outperforms most of the DEM-derived landslide controlling factors.
- Mahalanobis distance improves the accuracy and prediction power of MaxEnt.

The aim of the current study is to map landslide susceptibility over the Ziarat watershed in the Golestan Province, Iran, using Maximum Entropy (ME), as a machine learning model, with two sampling strategies: Mahalanobis distance (MEMD) and random sampling (MERS). To this aim, a total of 92 landslides in the watershed were recorded as point features using a GPS (Global Positioning System) device, along with several field surveys and available local data. By reviewing landslide-related studies and using principal component analysis, 12 landslide-controlling factors were chosen namely altitude, slope percent, slope aspect, lithological formations, proximity (to faults, streams, and roads), land use/cover, precipitation, plan and profile curvature and the state-of-the-art topo-hydrological factor known as height above the nearest drainage (HAND). Two sampling methods were used to divide landslides into two sets of training (70%) and test (30%). The Area under the success rate curve (AUSRC) and the area under the prediction rate curve (AUPRC) were used to evaluate the results of the MEMD and MERS. The results showed that both MEMD and MERS strategies with the respective AUSRC values of 0.884 and 0.878, have good performance in modelling the landslide susceptibility in the study area. However, AUPRC test showed slightly different results in which MEMD with the value of 0.906 showed excellent predictive power in comparison with the MERS with the AUPRC value of 0.846. The higher AUPRC value in relation to AUSRC indicated the MEMD as the premier model in the current study. According to the MEMD, three landslide controlling factors including lithological formations, proximity to roads and precipitation with the respective contribution percentages of 25.1%, 23.3%, and 19.1%, contained more information in relation to the rest. Moreover, according to one-by-one factor removal test, lithological formations and proximity to faults were identified to have a unique information compared to the rest. According to the MEMD, about 13.8% of the study area is located within high to very high susceptibility classes which can be matter of great interest to decision makers and the local authorities for formulating land use planning strategies and implementing pragmatic measures.