Prediction of Design Discharge Values in Ungauged Catchments in SloveniaAl Amin, Rafsan (Avtor) Bezak, Nejc (Mentor) Šraj, Mojca (Komentor) master thesisflood risk managementdesign discharge predictionmulti-variate regressionartificial intelligencegeo-statisticstopological krigingtabPFNungauged catchmentsReliable design discharge estimates are essential for flood risk management and hydraulic design, yet their estimation remains challenging in ungauged catchments where long-term discharge observations are unavailable. This study compared regression-based, artificial intelligence, and geostatistical regionalization methods for predicting design discharge in Slovenia using morphometric, terrain, land-cover, climatic, and rainfall-related catchment characteristics. Different target formulations were evaluated using leave-one-out cross validation, and the better-performing approaches were subsequently applied to an additional set of catchments treated as ungauged. Catchment size emerged as the dominant control variable on design discharge, with forest area repeatedly identified as an informative predictor because of its close relationship with catchment scale. Regression methods provided useful and interpretable predictions, while artificial intelligence and geostatistical methods generally offered greater flexibility in representing complex relationships. TabPFN showed particularly consistent performance across different target formulations and predictor combinations, while Top-Kriging demonstrated that strong predictions could also be obtained using relatively small external drift sets. Increasing the number of predictors did not consistently improve performance, showing the importance of parsimonious model development. Application to the ungauged catchments showed that no single tested method can be considered universally preferable. Model suitability depends on the characteristics of the target catchment, predictor availability, target formulation, model complexity, uncertainty, and the degree of extrapolation beyond the training data. The results therefore support a method selection approach in which several regionalization techniques are evaluated and the most appropriate model is chosen according to the specific ungauged catchment and available information.[R. Al Amin ]20262026-09-01 08:48:10Magistrsko delo/naloga186404sl