Two statistics for evaluating parameter identifiability and error reduction

Two statistics are presented that can be used to rank input parameters utilized by a model in terms of their relative identifiability based on a given or possible future calibration dataset. Identifiability is defined here as the capability of model calibration to constrain parameters used by a model. Both statistics require that the sensitivity of each model parameter be calculated for each model output for which there are actual or presumed field measurements. Singular value decomposition (SVD) of the weighted sensitivity matrix is then undertaken to quantify the relation between the parameters and observations that, in turn, allows selection of calibration solution and null spaces spanned by unit orthogonal vectors. The first statistic presented, “parameter identifiability”, is quantitatively defined as the direction cosine between a parameter and its projection onto the calibration solution space. This varies between zero and one, with zero indicating complete non-identifiability and one indicating complete identifiability. The second statistic, “relative error reduction”, indicates the extent to which the calibration process reduces error in estimation of a parameter from its pre-calibration level where its value must be assigned purely on the basis of prior expert knowledge. This is more sophisticated than identifiability, in that it takes greater account of the noise associated with the calibration dataset. Like identifiability, it has a maximum value of one (which can only be achieved if there is no measurement noise). Conceptually it can fall to zero; and even below zero if a calibration problem is poorly posed. An example, based on a coupled groundwater/surface-water model, is included that demonstrates the utility of the statistics.