Temporal and spatial transferabilities of hydrological models under different climates and underlying surface conditions

Highlights

• An evaluation method of model transferability is proposed.

• The complex models (XAJ and HBV) have better spatial transferability than the simple model (GR4J).

• The performance of models transferred from the wettest to the driest period is worse than that of models transferred from the driest to the wettest period.

Abstract

Changing conditions of the climate and underlying surface have altered the rainfall-runoff relationships in many basins, greatly increasing additional challenges in the applicability of hydrological models for studying the hydrological response to those potential changes. However, systematic and simultaneous testing and comparing of both temporal and spatial transferabilities of different hydrological models under changing conditions have not received enough attention. The present study investigates the potential differences between temporal and spatial transferabilities of different hydrological models under different climatic and underlying surface conditions, which are synthesized from two basins in Southern China with 50-year historical records (1966–2015). The transferability of five hydrological models, i.e., XAJ, HBV, SIMHYD, IHACRES and GR4J, is investigated under the synthesised changing conditions by using a new evaluation method, proposed in this study. The results show that: (1) the proposed evaluation method is proved to be effective in evaluating the transferability of the models; (2) for temporal transferability under stationary condition, the five models show similar performances, but for spatial transferability, the performances of complex models (XAJ and HBV) are better than that of the simple model (GR4J); (3) the difference in underlying surface conditions in the target basin affects spatial transferability of the models; (4) hydrological models have much better transferability from dry to wet period than otherwise. This study provides an insight to test temporal and spatial transferabilities of hydrological models in the context of changing climate and underlying surface conditions.

Introduction

The global climate and land use changes caused by substantial anthropogenic activities affect regional rainfall-runoff relationships, directly affecting local water resource availability (Arnell, 2004, Frich et al., 2002, Lu and Qin, 2020, Ma et al., 2008, Ragettli et al., 2020, Ye et al., 2013, Zhang et al., 2011, Zhang et al., 2012). Scientific and accurate assessments of future water resources under changing environment have attracted more attention than before because water-related issues, such as flooding, drought and pollution, are becoming increasingly grave due to the impact of global warming and human activities (Alcamo et al., 2007, Chen et al., 2019b, Döll, 2002, Li et al., 2015, Milly et al., 2008, Xiong et al., 2019). Hydrological models are the most important tool to study the impact of the changing environment on water resources (Chen et al., 2019b, Fan et al., 2019, Guo et al., 2019, Xu and Singh, 2004). Hydrological models have several advantages in studying the impact of environment change (Gleick, 1986, Jiang et al., 2007, Klemes, 1986, Schulze, 1997). Firstly, many models are already available for different climatic or physiographic conditions, increasing flexibility in identifying and choosing the most appropriate model to evaluate any specific region. Secondly, extensive climate change scenarios obtained by climate models can be used as inputs for hydrological models when assessing the hydrological response to climate change. Thirdly, hydrological models are easy to manipulate and improve for specific areas or conditions. They are usually calibrated by using historic records, assuming that conditions of the model application period will be similar to those of the calibration period (Jiang et al., 2007, Xu, 1999, Xu et al., 2005). However, altered rainfall-runoff relationship caused by climate and land use changes has also created some limitations and challenges in the use of hydrological models, which may cause the established models to become less skillful or lose their prediction ability in the new environment (Klemes, 1986). Therefore, it is essential to study the transferability of hydrological models in a changing environment.

Many studies on testing the applicability of hydrological models in changing climatic conditions have shown that many models do not have good temporal transferability, especially under non-stationary climatic conditions (Boorman and Sefton, 1997, Cornelissen et al., 2013, Eregno et al., 2013, Jiang et al., 2007, Li et al., 2012, Merz et al., 2011, Panagoulia and Dimou, 1997). Moreover, the studies revealed that different hydrological models delivered different results when simulating hydrological responses to future climate change scenarios. Coron et al. (2012) used three lumped models (GR4J, MORDOR6 and SIMHYD) to simulate runoff processes in 216 watersheds in south-eastern Australia and found that the greater the climate difference between the calibration and validation periods, the worse was the transferability of the models. Broderick et al. (2016) used six lumped hydrological models to conduct a cross-validation study by dividing dry and wet years in the 37 watersheds of Ireland; results showed that model transferability depended on the selected catchment, tested scenarios and evaluation criteria. Oni et al. (2016) used historical wet and dry years as a proxy for expected future extreme conditions in a boreal catchment, demonstrating that runoff may be underestimated by at least 35% when model parameters were transferred from dry to wet years.

Hydrological models’ spatial transferability has been studied using regionalisation methods (Bao et al., 2012, Merz and Blöschl, 2004, Parajka et al., 2013, Samuel et al., 2011, Swain and Patra, 2017, Yang et al., 2017, Yang et al., 2019, Yang et al., 2020). Yang et al. (2019) applied a lumped conceptual hydrological model (WASMOD) to investigate the transferability of regionalisation methods under changing climate conditions, based on 108 catchments in Norway. Lute and Luce (2017) built snow models of varying complexity in the western U.S. to evaluate model transferability in new locations and periods, indicating that the transferred models performed well in the new location with conditions similar to the trained location. They also found that simple to moderately complex models performed better than complex models when transferred to new locations in their study. Different results are reported by Yang et al. (2020) who tested spatial transferability of five conceptual hydrological models with varying number of parameters from 6 to 17, and concluded that the model with more parameters produced better results in most cases. A comprehensive survey of literature shows that there is no consistent conclusion about which regionalisation method or model performs best. Moreover, climate conditions are changing or are becoming non-stationary (IPCC, 2014), and under non-stationary climate conditions, the reliability of the model’s spatial transferability needs to be investigated. Therefore, it is very meaningful to further jointly study temporal and spatial transferabilities of different hydrological models under different climatic periods and in different basins.

The problem of general model transferability (spatial and temporal) has been recognised early as the major aim and the most difficult aspect of hydrological modelling (Klemes, 1986, Xu, 1999). Despite this fact, less attention has been paid to the testing of this most important aspect, compared with many other modelling issues like manual versus automatic calibration, optimisation, regionalisation, etc. (Klemes, 1986, Xu, 1999). In other words, operational testing of the models is not given the priority it deserves. Xu (1999b) made a preliminary attempt to evaluate temporal and spatial transferabilities of a lumped model in different simulation strategies; however, the study was limited by the number of models and data available at that time.

Above discussion reveals that although previous studies have explored transferability of hydrological models, some key issues are yet to be studied, which motivated the current study: (1) How do temporal and spatial transferabilities of hydrological models differ with the model complexity? (2) How do temporal and spatial transferabilities of hydrological models depend on different climates and underlying surface conditions of the basin? (3) What are the performance differences when the models are calibrated under dry/wet condition and transferred to wet/dry condition? To achieve these goals, five lumped hydrological models, including XAJ (Zhao et al., 1980), HBV (Bergstrom, 1976), SIMHYD (Chiew et al., 2002), IHACRES (Jakeman et al., 1990), and GR4J models (Perrin et al., 2003) with different complexities and flow generation methods are applied to two catchments in central-south China in this study. The temporal and spatial transferabilities of the five conceptual models are compared and analysed by using the split-sample, differential split-sample, proxy-basin and differential proxy-basin tests under stationary and changing conditions, including different climatic periods, different basins and their combinations. The rest of this paper is organised as follows. Section 2 introduces the study area and data. Section 3 provides the details about the five lumped models, and model calibration and validation methods. Section 4 presents and discusses the results corresponding to different simulation strategies. Finally, Section 5 draws major conclusions and presents the limitations and possible future development of this study.