Improving Landslide Susceptibility Mapping via Non-Landslide Sampling Strategies

This study assesses the impact of non-landslide sample selection on landslide susceptibility mapping. Two strategies were compared: (i) randomly selecting negatives outside buffers around mapped landslides (S1) and (ii) a targeted, threshold-and-buffer method (S2). A dataset from Sulaymaniyah Governorate, Iraq, including 148 landslides and 434 non-landslide points, was modeled using Logistic Regression (LG) with 14 conditioning factors derived in ArcGIS. Factors were ranked using the Frequency Ratio (FR). The five most influential ones—slope, Topographic Wetness Index (TWI), soil, Normalized Difference Vegetation Index (NDVI), and Land Use Land Cover (LULC)—were binary-reclassified to delineate safe zones. Additionally, S2-negative samples were collected within a 500–750 m annulus. Performance was evaluated using confusion matrices and Receiver Operating Characteristic – Area Under the Curve (ROC–AUC) on a 75/25 split. S2 achieved accuracy of 89.6%, precision of 77.5%, and AUC of 93.7%, outperforming S1 (85.4%, 70.0%, 91.2%). When validation was limited to landslide points, S1 exhibited a slightly higher AUC (88.4% vs. 85.3%), indicating greater sensitivity but lower precision. Results show that threshold-guided, proximity-constrained negatives enhance class separation and lower false positives without altering the learning algorithm, supporting their application for more reliable susceptibility mapping.