Accurately predicting runoff responses in ungauged catchments remains a central challenge in hydrology, particularly for regions lacking sufficient observational data. The regionalization approach is a powerful strategy to address this challenge, enabling the extrapolation of knowledge derived from gauged catchments to ungauged locations. However, the development of robust regionalization models is often constrained by limited datasets, in terms of the number of catchments and/or the availability of catchment attributes. Despite recent advancements, most existing studies evaluate model performance using subsets of the same dataset for both training and testing, limiting the assessment of model transferability. To address this gap, this study introduces a regionalization framework using the large-scale CAMELS-US dataset to train a set of regression models for predicting the calibrated parameters of the GR2M, a monthly water balance model. The framework is then evaluated on the CAMELS-GB (Great Britain dataset), representing a distinct climatic and geographic context beyond continental boundaries. A critical aspect of this study is its focus on uncovering the physical meaning of the GR2M parameters by linking them to measurable catchment attributes. Results show that while linear regression models perform well within the training domain, their predictive skill declines significantly when applied to the GB dataset. In contrast, machine learning models, particularly Support Vector Regression, demonstrate strong generalization capabilities, achieving high accuracy in predicting runoff responses across GB catchments. These findings underscore the importance of nonlinear approaches in developing globally transferable regionalization frameworks and underscore their potential to support reliable hydrologic predictions in data-scarce regions.
扫码关注我们
求助内容:
应助结果提醒方式:
