Estimating a geotechnical factor of safety: A Python-based workflow for regional landslide susceptibility assessment

Abstract

Background: Assessing slope stability using geotechnical engineering principles typically uses a Factor of Safety (FS) to define the ratio of resisting to driving forces along a potential failure surface of landslides. Accurate estimation of this FS is central to evaluating shallow landslide susceptibility, yet field-based measurements of soil strength are rarely available across steep, remote terrain. This challenge is particularly evident in New Zealand’s erodible hill-country. To address this limitation, this study developed a reproducible Python-based workflow for computing FS using environmental geographic information system (GIS) layers and published geotechnical parameters, without requiring in-situ testing.

Methods: To demonstrate the efficacy of the method, a case study area was selected in Gisborne, New Zealand. In this region, major storms triggered numerous shallow landslides in a plantation forest area that was later accurately mapped by forest owners. Geological and soil units were assigned representative shear-strength properties using values compiled from national datasets and previous studies. Terrain variables, including slope, curvature, elevation, drainage metrics, soil attributes, and rainfall, were extracted at 42,000 landslide and non-landslide sample locations. Landslide samples were derived from the centroids of 22,000 mapped shallow landslide polygons, representing about 4,400 ha of documented landslide-affected terrain. Non-landslide samples were generated using a terrain-constrained random sampling approach within the same study area, explicitly excluding mapped landslide polygons and non-terrestrial areas. An infinite-slope model was implemented in Python to compute FS under a range of hydrological and soil-depth scenarios.

Results: For this case study area, the median FS across the dataset was 1.3, and approximately 28% of samples fell below the instability threshold (FS < 1). Sensitivity analysis confirmed that FS was strongly responsive to increasing saturation ratio and moderately affected by soil depth. A qualitative comparison with mapped landslides showed that unstable and near-critical conditions were more frequently associated with inventoried failure locations, with 37% of landslide samples exhibiting FS < 1 compared to 18% of non-landslide reference points.

Conclusions: The results demonstrate that a transparent, Python-based workflow can generate physically interpretable FS estimates suitable for regional-scale screening in data-limited environments. The workflow establishes a foundation for future hybrid models in which FS-derived indicators can support, constrain, or enhance machine-learning frameworks of landslide susceptibility.