نوع مقاله : پژوهشی
نویسندگان
1 دانشکده مهندسی عمران، دانشگاه صنعتی شریف
2 دانشکده مهندسی عمران و محیط زیست، دانشگاه UNSW، سیدنی، استرالیا
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Global sensitivity analysis is a key component of hydrological modeling, enabling quantification of how changes in input variables influence simulated outcomes and supporting model development, calibration, and decision-making. Conventional sampling-based approaches, such as variance-based global sensitivity analysis, have been widely applied, yet they require substantial computational effort and very large numbers of model runs. These challenges have encouraged the development of data-driven methods that rely on available data, reduce computational burden, and avoid the need for predefined sampling schemes, making them an appealing alternative in many practical modeling situations where computational resources are limited. This study compares three data-driven global sensitivity analysis algorithms: permutation importance, partial dependence, and Friedman’s H-statistic. Each algorithm is described in terms of its underlying principles, advantages, and limitations for modeling variable importance and interactions. To evaluate their performance, the Sobol-G function, a widely used benchmark model, and the HBV rainfall–runoff model, a representative conceptual hydrological model, were employed, allowing assessment across both controlled mathematical settings and real-world hydrological conditions. The results indicate that permutation importance, when used with machine learning models such as random forests, typically obtains accurate ranking of influential variables and effective characterization of interaction effects, particularly in complex, nonlinear, and high-dimensional problems. These characteristics are especially important in hydrological applications, where identifying dominant drivers of model behavior is essential for reliable forecasting, uncertainty reduction, and system understanding. In addition to numerical measures, the role of visualization in global sensitivity analysis is highlighted. Visual tools such as bar charts, heatmaps, network diagrams, and partial dependence plots are described and compared, illustrating how they enhance interpretation and communication of findings by providing intuitive summaries of variable effects and interactions. However, caution is advised against over-reliance on visual representations without careful contextual examination, as misleading patterns may appear when graphs are misinterpreted or taken at face value. Overall, the study underscores the significance of data-driven sensitivity analysis as a flexible, efficient, and interpretable approach for improving hydrological modeling and decision support under uncertainty, particularly when computational constraints or model complexity limit the use of traditional sampling-based methods.
کلیدواژهها [English]
)