تجزیه و تحلیل حساسیت از توزیع مدل های شبیه سازی محیط زیست: درک رفتار مدل در مطالعات هیدرولوژیکی در مقیاس حوضه آبریز

The development of new hydrological simulation tools allows for the modelling of large hydrological catchments, with the aim of comprehensive management of the water resources, control of diffuse pollution processes, such as the fate of agricultural fertilizants and finally, with purposes of economical optimization of the crop yields as a function of the expected climate, the watershed characteristics and the socio-economical conditions of the region where the catchment is located. This paper describes the sensitivity analysis of a hydrological distributed model applied in one large European watershed by using a two-step procedure. Firstly, it allows for the consideration of a huge input parameter data set by using an implementation of the Morris screening procedure, eschewing the huge computational requirements arising from the necessary repetitive simulations. In the second step it provides quantitative estimations of sensitivity in terms of variance decomposition procedures based upon the FAST method for both the hydrological and the water quality determinants.

In recent years, water management became an important issue increasing the need of enhancing existent hydrological models. Major advances on this issue include the development of codes allowing the distributed simulation at the catchment scale on a daily basis, accounting for the hydrological and the water quality processes in both the inland and the channels. Usually the calibration and validation of such models require a huge amount of data, including both in situ and laboratory measurements as well as good quality meteorological records. The implementation of sensitivity analysis procedures is a useful tool in the calibration of the models and also in their transposition to different watersheds. However, the huge number of parameters becomes a drawback with respect to the capabilities of the sensitivity analysis methods, whose computational requirements increase noticeably with the number of parameters analysed. The present paper shows a two-step procedure allowing for the analysis of complex distributed hydrological models. The first step consists of a screening procedure leading to a qualitative ranking of the whole set of input parameters for different model outputs with relatively low computational cost. During the second step, a Fourier amplitude sensitivity test (FAST) technique is applied to the most relevant parameters for a set of specific model output obtaining quantitative measures of sensitivity in terms of their contribution to the output variance. A theoretical background of both the hydrological model and the sensitivity analysis methods is briefly introduced in Section 2. Section 3 is devoted to describe the methodology proposed in the paper from an algorithmic point of view, and focusing especially in the preliminary sampling of the input parameters. The description of the investigated watershed is presented in Section 4.1. Several important aspects are also discussed along with: the influence of the desegregation scheme of the basin is commented in Section 4.2; in Section 4.3 some results from the calibration of the model in a monthly basis are introduced. Regarding the input/output sets, the probability density functions (PDFs) adopted for the set of parameters and the list of investigated variables are enumerated in 4.4 and 4.5, respectively. The results of this work are described in detail in Section 5. The qualitative ranking obtained in the first step, its convergence and its analysis are, respectively, described in 5.1, 5.2 and 5.3, whereas the FAST quantitative estimates are summarized in Section 5.4. Finally, some conclusions and lines of further research are introduced in Section 6.

The proposed two-step methodology allowed for the sensitivity analysis of complex hydrological models on large watersheds. The screening step (Morris) allowed for the analysis of the whole data set without considering strong initial hypothesis on the input parameters. The ranking produced in this step helped the retrieval of the most sensitive inputs affecting each specific determinant. The FAST procedure provided quantitative measures of sensitivity based on such a parameter selection. In general, the ranking produced by the Morris method agreed with the FAST sensitivity estimations. However, some experimental deviations were also drawn, mainly due to the absence of some of the relevant parameters with respect to several output variables when performing the FAST estimations. Performing a specific FAST exercise for each output variable with its 14 more important parameters (i.e. if the computational cost is affordable) may minimize such deviations but it was not possible in this case study. In this research three FAST exercises were defined to address the analysis of 22 output variables. It was demonstrated that even if the relevant parameter subset retrieved from the Morris ranking was not the best possible, the FAST variance decompositions of most of the output variables was acceptable in many cases (Fig. 12). In general the variance of the model outcome was acceptably explained (e.g. 85–100%) by the variance of a few parameters. In the Ouse River Basin the calibration was performed manually before of carrying out the sensitivity analysis. However, the results obtained prove that this analysis might have been useful also in the calibration step. Another positive conclusion of this work is to point out the relevance of the soil parameters and hence the retrieval of comprehensive soil measurements in order to obtain better implementations of the model. In general, soil properties resulted to be much more important than crop and management parameters on the studied output variables. It is also important to remark that some SWAT-specific parameters (e.g. ESCMPC, REVAPC, NPERCO) showed high impacts on several important outputs (e.g. the surface water cycle, return water flow and nitrate runoff, respectively). Their degree of representativity under different catchment conditions should be investigated deeper. This study was performed in the basis of annual averages over the simulation period and might be improved by analysing the daily or monthly values. In addition, the enhancement of the screening procedure in order to distinguish the non-linear and crossed effects among the input data set is also a further line of research.