Research output: Contribution to Journal/Magazine › Journal article › peer-review
Research output: Contribution to Journal/Magazine › Journal article › peer-review
}
TY - JOUR
T1 - Sensitivity analysis based on regional splits and regression trees (SARS-RT)
AU - Pappenberger, F.
AU - Iorgulescu, I.
AU - Beven, Keith J.
PY - 2006/7
Y1 - 2006/7
N2 - A global sensitivity analysis with regional properties is introduced. This method is demonstrated on two synthetic and one hydraulic example. It can be shown that an uncertainty analysis based on one-dimensional scatter plots and correlation analyses such as the Spearman Rank Correlation coefficient can lead to misinterpretations of any model results. The method which has been proposed in this paper is based on multiple regression trees (so called Random Forests). The splits at each node of the regression tree are sampled from a probability distribution. Several criteria are enforced at each level of splitting to ensure positive information gain and also to distinguish between behavioural and non-behavioural model representations. The latter distinction is applied in the generalized likelihood uncertainty estimation (GLUE) and regional sensitivity analysis (RSA) framework to analyse model results and is used here to derive regression tree (model) structures. Two methods of sensitivity analysis are used: in the first method the total information gain achieved by each parameter is evaluated. In the second method parameters and parameter sets are permuted and an error rate computed. This error rate is compared to values without permutation. This latter method allows the evaluation of the sensitivity of parameter combinations and thus gives an insight into the structure of the response surface. The examples demonstrate the capability of this methodology and stress the importance of the application of sensitivity analysis.
AB - A global sensitivity analysis with regional properties is introduced. This method is demonstrated on two synthetic and one hydraulic example. It can be shown that an uncertainty analysis based on one-dimensional scatter plots and correlation analyses such as the Spearman Rank Correlation coefficient can lead to misinterpretations of any model results. The method which has been proposed in this paper is based on multiple regression trees (so called Random Forests). The splits at each node of the regression tree are sampled from a probability distribution. Several criteria are enforced at each level of splitting to ensure positive information gain and also to distinguish between behavioural and non-behavioural model representations. The latter distinction is applied in the generalized likelihood uncertainty estimation (GLUE) and regional sensitivity analysis (RSA) framework to analyse model results and is used here to derive regression tree (model) structures. Two methods of sensitivity analysis are used: in the first method the total information gain achieved by each parameter is evaluated. In the second method parameters and parameter sets are permuted and an error rate computed. This error rate is compared to values without permutation. This latter method allows the evaluation of the sensitivity of parameter combinations and thus gives an insight into the structure of the response surface. The examples demonstrate the capability of this methodology and stress the importance of the application of sensitivity analysis.
KW - Regression tree
KW - Sensitivity analysis
KW - Random Forests
KW - Uncertainty analysis
KW - Calibration
KW - Generalized likelihood uncertainty estimation
KW - Regional sensitivity analysis
U2 - 10.1016/j.envsoft.2005.04.010
DO - 10.1016/j.envsoft.2005.04.010
M3 - Journal article
VL - 21
SP - 976
EP - 990
JO - Environmental Modelling and Software
JF - Environmental Modelling and Software
SN - 1364-8152
IS - 7
ER -