Structural Improvement And Nonstationarity Of Hydrological Model From Aspect Of Evapotranspiration

Watershed hydrological models play an important role in hydrological forecasting and water resources management.For the model spatial transferability,remotely sensed actual evapotranspiration（RS-ET）products have been widely used for hydrological modeling in ungauged basins.However,the existing approaches of using RS-ET data can only improve the accuracy of runoff simulation for minority models/minority basins and is lack of universality.For the model temporal transferability,fix ET structues are generally used for hydrological modeling.However,ET processes may change under climate change and human activities,therefore the assumption that the hydrological model structure is time-invariant is no longer appropriate.In this study,basins in the United States from the MOPEX datasets and basins suffered from the Millennium drought in Australia are selected as the study regions.Focused on catchment evapotranspiration process,the conventional approaches of implementing RS-ET data have been modified to improve the model spatial transferability and further to improve runoff simulation in ungauged basins.Furthermore,time-variant model structures considering the nonstationarity of hydrological processes have been adopted to improve the model temporal transferability and further to improve runoff simulation in the changing environment.The main contents are summarized as follows:（1）For the application of RS-ET data in hydrological models,two approaches have been proposed based on the evaluation of ET simulation efficiency and assessment of RS-ET data accuracy.Taking account of compatibility between RS-ET data and different hydrological model structures,one approach incorporates RS-ET into hydrological models as direct input only under the premise that the ET simulation efficiency exceeds the NSEthr.Furthermore,considering the fact that compared with rainfall periods,the uncertainty in ET simulation of hydrological models can be larger while and the accuracy of RS-ET observations can be higher duiring rainless periods,another approach utilizes RS-ET for model multiobjective calibration only during rainless periods,when ET simulation in hydrological models contains larger uncertainty and RS-ET data are of relatively higher accuracy.An intercomparison among these modeling approaches is conducted over 208 MOPEX basins in the United States using three hydrological models（Xinanjiang model,SIMHYD and GR4J）and RS-ET data retrieved from AVHRR.Then,the hydrological processes in ungauged basins are modeled using the optimized parameter values from the geographically nearest gauged basins.Results show that: Compared with the conventional modeling approach,using RS-ET as partial direct input only under the premise that NSEE > NSEthr improves streamflow simulation efficiency in 91.1% basins when using the Xinanjiang model,and the NSEQ can be improved more than20% in basins with annual precipication below 1200 mm and originally poorly simulated basins（NSEQ < 50%）;the approach of using RS-ET for multiobjective calibration only during rainless periods significantly improves both NSEQ and NSEE in regionalization by 5.3% and14.2% respectively,and the improvement of NSEQ and NSEE can exceed 10% and 15%respectively in originally poorly simulated basins（NSEQ < 60% and NSEE < 50%）.（2）For the potential change of ET physical process under environment change,a timevariant hydrological model structure has been proposed based on the nonstationarity of ET process.Firstly,five hydrological models containing different ET structures（GR4J,GR4J_Budyko,GR4J_sequential,GR4J_weighting and GR4J₃E）are used for hydrological modeling over 208 MOPEX basins in the United States,then an intercomparison among the results of five models is conducted over differnet periods,finding that relative performances of different ET structures vary in time.On this basis,these five models are averaged using the BMA method,and a time-variant ET structure is constructed by optimizing the BMA weights over different sliding windows.Results show that: Compared with the best single model structure,the adoptation of time-variant ET structure has improved NSE[ Q] by14%.（3）For the structural nonstationarity of hydrological models under environment change,a BMA based method has been proposed for identifying the time-variant model structure.Firstly,multiple model structures are averaged using the BMA method and the BMA weights are optimized over different periods and in different sliding windows,threrefore the model structural nonstationarity can be translated into the time-variant BMA weights by evaluating the temporal variations of weights.A time-variant model structure is constructed by using the Xinanjiang and GR4 J model in six basins with significant climate change.Results show that:The temporal variations of hydrological model structure can be interpreted from the perspective of climate change,and specifically,the weight of Xinanjiang model increases with larger precipitation,and vice versa;compared with fixed model structure,the adoptation of a time-variant model structure has improved NSE[ Q] by 19%.