Landslide susceptibility prediction using the coupled Mahalanobis distance and machine learning models (case study: Owghan watershed, Golestan province)

Landslide susceptibility maps are considered a backbone for decision-makers to suggest solitary or combined technical and regulatory measures. Such maps are also considered an invaluable tool for engineers, earth scientists, planners, and decision-makers to select the most suitable areas for agriculture, building, and other development activities. Hence, thanks to landslide susceptibility maps, addressing highly susceptible areas are feasible, so that over the course of further detailed studies on the imminent landslide occurrences in the future, landslide potential risk is mitigated.

In this study, two robust data mining models, namely random forest and maximum entropy were used to map landslide susceptibility across the Owghan Watershed in Golestan province. After preparing the landslide inventory map via extensive field surveys, interpreting Google Earth images, and the archived data acquired from different organizations, landslide points were split into two sets of training (70%) and validation (30%) by using the Mahalanobis distance technique. Further, drawing on the extensive literature review, fifteen factors including climatic, geological, tectonic, topo-hydrological, and anthropogenic drivers, as landslide-controlling factors were selected and sieved through the variance inflation factor test. Ultimately, by implementing the above-mentioned data mining techniques, the most important factors in the modeling process, as well as the highly susceptible locations in the study area, were introduced.

Evaluating the learning capability, both the random forest and maximum entropy models with the respective area under the receiver operating characteristic curve (AUROC) values of 0.923 and 0.91, showed almost identical fitting abilities. However, getting to the validation stage, it was found that the random forest with the AUROC value of 0.9 clearly outperforms maximum entropy (AUROC= 0.85) in terms of prediction power and generalization capacity. Hence, the random forest was suggested as a better-performing model for landslide susceptibility mapping in the Owghan watershed, compared to its counterpart. About 10% of the study area falls into high and very high landslide susceptibility zones. Furthermore, five landslide-controlling factors including rainfall, normalized difference vegetation index, height above the nearest drainage, lithological formation, and proximity to roads have been found to be the most significant factors contributing to landslide occurrence in the study area. Additionally, the results attest that announcing the Safiabad village as a landslide-prone area by the authorities is technically sound and evacuating the residents to a new place has been a right decision; however, some parts of the newly inhabited area shows landslide predisposing patterns which can lead to a higher susceptibility of the area to landslide occurrence in the future.

Scrutinizing the results of random forest model revealed that a combination of natural factors (intense rainfall, bare lands, susceptible lithological formations, and topo-hydrological mechanisms) and anthropogenic interferences (tillage parallel to slope length/perpendicular to contour lines and unprincipled road construction) are synergistically responsible for landslide occurrence in the Owghan Watershed. On the other hand, announcing the Safiabad village as a critical landslide-prone area seems to be a wise decision, although the newly inhabited place seems to be selected merely based on having a suitable slope steepness (i.e., almost flat) and being accessible through several connecting routes, while the enhanced conservation tillage methods have not been applied to the selected site and adjacent areas. The latter, according to our inferences, can trigger a crisis in a larger extent. Moreover, owing to the presence of other landslide predisposing factors in the new residential site, safe areas should be pointed out and announced by adopting a holistic view on the entire influential and predisposing conditions for landslide occurrence.