Akademik Veri Yönetim Sistemi
13th INTERNATIONAL BILTEK CONGRESS ON CURRENT DEVELOPMENTS IN SCIENCE, TECHNOLOGY AND SOCIAL SCIENCES, Paris, Fransa, 18 - 21 Aralık 2025, ss.319-325, (Tam Metin Bildiri)
Accurate statistical downscaling methods are essential for translating large-scale climate datasets, such as Global Circulation Models (GCMs), into reliable local precipitation estimates. Among these methods, the Multivariate Adaptive Regression Splines (MARS) algorithm is widely used due to its ability to model nonlinear relationships using flexible basis functions. However, the internal mechanics of MARS can rely on two different modeling stages when forming model terms. The forward step generates a large set of basis functions by adding the most useful terms to the model, while the backward pruning removes redundant or lowcontribution terms. When used independently, the relative influence of each method on predictive accuracy can differ substantially in precipitation downscaling applications. This study investigates the performance differences between MARS models built using only the forward step and those refined through backward pruning, using monthly precipitation data. Both approaches were applied with identical predictor sets, and models were trained over 1979 2007 and tested over 2007 2014. Results show that forward-step-only models generally exhibit reduced performance during the test period. In contrast, backward pruning generally improved model performance according to the NSE metric. These findings highlight the crucial role of backward pruning in achieving accurate, stable, and practical MARS-based downscaling for climate and hydrological studies.