Accurate landslide susceptibility mapping (LSM) is essential for disaster risk reduction in seismically active mountain regions, yet the selection of mapping units remains a fundamental source of uncertainty in model outcomes. Although grid units (GU) are widely adopted due to their simplicity and compatibility with raster data, their implications for prediction stability, uncertainty behavior, and interpretability under co-seismic conditions remain insufficiently evaluated relative to slope units (SU). This study provides a systematic comparison of SU and GU frameworks for co-seismic LSM using the 2015 Mw 7.8 Gorkha earthquake in Nepal as a case study. Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) were implemented under both frameworks, with RF selected as the benchmark model and applied within a 20-run ensemble to explicitly assess prediction stability and uncertainty. Both mapping approaches achieved high predictive accuracy, with area under the curve values ranging from 0.90 to 0.95, demonstrating that conventional accuracy metrics alone do not distinguish mapping-unit performance. However, clear differences emerged in uncertainty behavior and spatial coherence: GU-based models exhibited higher sensitivity to sampling variability and greater dispersion of susceptibility probabilities, particularly within intermediate probability ranges (0.3–0.7), whereas SU-based models produced more spatially coherent susceptibility patterns with significantly reduced uncertainty. Aggregation of GU susceptibility outputs to the SU scale preserved dominant susceptibility signals, confirming the feasibility of unit transformation for comparative analysis. Overall, the results show that while GU-based modeling captures fine spatial detail, SU-based frameworks offer superior uncertainty control, geomorphological interpretability, and prediction stability, making them particularly suitable for co-seismic LSM applications where reliable hazard assessment and decision support are required.
Bhattarai et al. (Fri,) studied this question.