Browsing by Subject "support vector classifier"

Now showing 1 - 1 of 1

Access status: Open Access ,
Geospatial and Optimized SVM-Based Landslide Susceptibility Zonation of South District of Sikkim, India
(Wydawnictwa AGH, 2025) Anuragi, Saurabh Kumar
Landslide identification and susceptibility maps play vital roles in supporting planners and decision-makers who manage disaster risks. By providing accurate information, these maps significantly contribute to minimizing the potential losses of life and property. To create effective landslide-susceptibility models, it is essential to incorporate a combination of terrain characteristics and meteorological factors, thus enhancing our understanding and preparedness for such events. This study presents a comparative analysis of three kernel functions (linear, polynomial, and RBF) of an support vector classifier (SVC) accompanied by a grid-search in order to determine optimal hyper-parameter settings. The primary objective of this methodological framework is to ensure accurate and reliable predictions for the generation of landslide-susceptibility maps in the South District of Sikkim, India. In this investigation, 14 conditioning factors were considered, including aspect, distance to streams, distance to roads, drainage density, elevation, lithology, land use/land cover (LULC), normalized difference vegetation index (NDVI), plan curvature, profile curvature, rainfall, slope, soil type, and earthquake susceptibility. The performances of the models were evaluated using a range of metrics, including the training score, testing score, kappa, sensitivity, specificity, accuracy, and area under the curve (AUC). Optimal hyper-parameter tuning for each SVC kernel was conducted through a grid-search approach. The results indicated that the SVC_poly and SVC_rbf models surpassed the linear model, achieving accuracy and AUC values of 0.907 and 0.908, respectively, in developing susceptibility maps. Consequently, both the SVC_poly and SVC_rbf models were identified as the most reliable and effective tools for landslide-susceptibility mapping in this study, making them optimal choices for predictive analyses in this domain.