Optimizing slope unit-based landslide susceptibility mapping using the priority-flood flow direction algorithm

Regional-scale landslide susceptibility assessments are challenging and have recently captured the attention of scientists and technicians. A slope unit is a typical map unit that enhances the readability of landslide susceptibility maps (LSMs) by reducing ridge effects. Slope units are commonly obtained via the subbasin division (SBD) method from a digital elevation model (DEM) containing sink features, such as depressions, pits and flats. The SBD method addresses the depression problem by filling features but subsequently faces unresolved problems with parallel slope-unit boundaries. The main objective of this study was to introduce the priority-flood flow direction (PFD) algorithm to exploit flow information for depressions and compare PFD-improved slope units to SBD-based slope units in corresponding LSMs. An LSM case study was generated using an artificial neural network in Yan’an city, China, where loess ridge and hill landforms are developed. Data for a total of 447 landslides were obtained from historical records and field surveys to construct a landslide inventory map, where 313 landslides (70%) and 134 landslides (30%) were used for training and validation, respectively. Then, 9 conditioning factors related to landslides were prepared as inputs to generate six LSMs by combining two map units and three catchment area thresholds. These factors included elevation, slope, river distance, road distance, rainfall, normalized difference vegetation index (NDVI), aspect, land use and lithology. Finally, the resulting LSMs were validated and compared between the PFD and SBD methods using the precision, recall, accuracy, F1 score, success rate and prediction rate based on receiver operating characteristic (ROC) and cumulative landslide occurrence (CLO) curves. The results showed that the improved slope units produced no parallel boundaries, and the environment-factor layers generated based on these slope units prevented the generation of parallel factor-subclass boundaries, thus reducing striped distributions of factors and optimizing patterns in the LSMs. The PFD-improved LSM yielded ROC-based success rates ranging from 0.8849 to 0.8901, which were better than those obtained for the SBD-based LSM, at 0.8779–0.8823. The prediction rate revealed that the PFD-improved slope units were more suitable for large regions, as they provide more stable performance than the SBD-based slope units at large scales. The PFD algorithm resolved the problem of parallel boundaries and optimized LSM patterns, resulting in enhanced performance over the SBD method in large-scale research. The study results could be useful for local government management and land use planning in large regions.