Regionalization for Ungauged Catchments — Lessons Learned From a Comparative Large‐Sample Study

Abstract

Model parameter values for ungauged catchments can be regionalized from hydrologically similar gauged catchments. Achieving reliable and robust predictions in ungauged catchments by regionalization, however, is still a major challenge. Here, we conduct a comparative assessment of 19 regionalization approaches based on previously published literature to contribute new insights into their performance in different geographic regions. The approaches use geographical information, physical catchment attributes, hydrological signatures, or a combination thereof to select donor catchments and to subsequently transfer their entire parameter sets to the ungauged receiver catchment. Each regionalization approach was tested in a leave-one-out cross-validation with a bucket-type catchment model (the HBV model) using 671 gauged catchments in the United States with a diverse hydroclimatology. We then evaluated regionalization performance for several hydrograph aspects, compared it against calibration and regionalization benchmarks, and linked it to catchment descriptors. The results of this large-sample regionalization study can be summarized in three major lessons: (a) Catchments can benefit from a well-chosen regionalization approach independent of their geographic region and independent of how well they can be modeled or regionalized at best. (b) Almost perfect donors exist for most catchments and an excellent relative model performance can be reached for most catchments with current regionalization approaches. This implies that there is considerable potential for improvement in the prediction in ungauged catchments. (c) The ranking of regionalization approaches depends on how the predicted hydrographs are evaluated. These findings indicate that a multi-criteria evaluation is essential for a robust assessment of regionalization performance.Plain Language Summary Information on streamflow is crucial for good water resources management including the mitigation of water-related hazards. However, for many catchments there is a lack of streamflow information. In such situations, streamflow is often estimated using hydrological models, whereby model parameterizations are transferred (i.e., regionalized) from hydrologically similar gauged catchments. Reliable estimates in data-scarce regions are still a major challenge in hydrology despite the large number of regionalization approaches proposed in the past decades. Here, we conduct a systematic and standardized assessment of 19 existing regionalization approaches using 671 catchments in the United States. Our findings suggest that widely used regionalization approaches can result in excellent model performance for most catchments, whereby approaches considering spatial proximity and any kind of volume information are among the most promising ones. While volume information is per definition missing in ungauged catchments, it could possibly be derived from a small number of field measurements or estimated through statistical analysis. However, the most suitable approach can vary cons...

Prediction at Ungauged Catchments through Parameter Optimization and Uncertainty Estimation to Quantify the Regional Water Balance of the Ethiopian Rift Valley Lake Basin

“…Therefore, the 14 catchments produced a 14-regression model in the leave-one-out crossvalidation method that quantifies prediction uncertainty in the ungauged catchments. The leave-one-out spatial cross-validations also show better performance in the regionalization studies that use discharge signatures [15,18] and for ungauged catchments parameter prediction [7,8,65]. This method is more stable and more resilient to irreducible errors for large sample studies [63].…”

“…Other physical properties where local information is not available, such as permeability and porosity, were extracted from the global datasets prepared by Huscroft et al [44]. These dominant catchment properties are also widely applied in the previous regionalization studies using the HBV model structure [7,45,46].…”

“…For large sample training data, this method is more resilient to irreducible errors [62]. Recently, this technique is applied for regionalization studies involving the prediction of discharge signatures [15,18,63] and model parameters in the ungauged catchments [7,8,64]. Thus, for 14 catchments, there are 14 different regression models and 14 different evaluation catchments.…”

“…The Prediction for Ungauged Basins (PUB) initiative aims in particular at developing strategies for better understanding and reduced uncertainty in data-scarce regions [3,4]. Studies have started to focus on ungauged catchment predictions during the PUB decade [4,5] and in the post-PUB decade [6], including those comparing existing regionalization approaches for large sample studies [7]. Two general approaches have been used for predictions in ungauged basins: the first estimates model parameters from calibrated model parameters based on selected objective functions [8][9][10][11]; the second is a model-independent approach, which uses streamflow signatures to establish constraints that can describe the physical and climatic characteristics of watersheds [12].…”

“…Most model parameters are not directly measurable or linkable to the physical properties of the given catchment because of model simplifications or disagreements between the model scale and the observation scale (incommensurability) [21][22][23]. However, an inherent correlation between the model parameters and catchment properties can often be assumed [5,7,24,25]. Model parameters represent the characteristics of the complex catchment system that are difficult to measure on a small scale.…”