COMPARATIVE ANALYSIS OF SPATIAL LANDSLIDE SUSCEPTIBILITY MODELS USING THE SZ PLUGIN IN QGIS, UNDER DATA-LIMITED CONDITIONS. CASE STUDY: SĂCELE RESERVOIR IN BRAȘOV, ROMANIA
Sorin-Gabriel VATAMANU1
1 Babeș-Bolyai University, Faculty of Environmental Sciences and Engineering, , 30 Fântânele Str., 400294 Cluj-Napoca, Romania, e-mail: sorin.vatamanu@stud.ubbcluj.ro.
ABSTRACT. – Comparative Analysis of Spatial Landslide Susceptibility Models Using the SPATIAL-ZONING Plugin in QGIS, Under Data-Limited Conditions. Case Study: Săcele Reservoir in Brașov, Romania. Landslide susceptibility in data limited environments represents a methodological challenge. This study investigates spatial landslide susceptibility at Săcele reservoir in Brașov, Romania, by using an open-data reproducible approach implemented by the use of the Spatial-Zoning plugin in QGIS. Slope units were used as the primary mapping units for modeling, where independent parameters were derived freely from available geospatial datasets by using an extraction script in GEE (Google Earth Engine). The landslide inventory was created through in situ survey and satellite image analysis, identifying a total of 33 landslides within the study area. In order to evaluate robustness under small sample and imbalanced conditions, three approaches were tested, a) machine learning (ML) algorithms such as: Random Forest (RF), Support Vector Machine (SVM), and b) statistical tools as Generalized Additive Model (GAM). Firstly, model fittings were performed with k fold = 1, after which 5 fold cross-validation yielded comparable mean AUC values for RF and SV, (≈0.70), but with substantial variability (RF: 0.50-0.93, SVM: 0.52-0.86). GAM achieved a higher and more stable performance (AUC≈0.75, range 0.69-0.85), indicating superior generalization capacity. Considering predictive accuracy
and stability, the GAM model was selected to produce the final susceptibility index map by using a confusion matrix. The findings of this study demonstrate that semiparametric models may outperform more complex machine-learning approaches
in small sample susceptibility studies.
Keywords: GIS, Spatial-Zoning plugin, landslide inventory, machine learning, generalized additive model, AUC, slope units