E-WP6 - DroughtClip

Project aims

Climate model output and therefore climate predictions exhibit systematic departures from the observed climate state, especially on small regional scales. This bias reduces prediction skill and restricts usability and value of the predicted quantities. These deviations however, can often be corrected with statistical post-processing methods to obtain more meaningful results. Bias and its correction for skewed variables like wind speed and precipitation is the subject of DroughtClip during the second phase of MiKlip. The main goals are:

  • Develop improved bias correction methods for the MiKlip decadal prediction system
  • Investigate their impact for user-relevant parameters such as droughts with the aim to improve prediction skill
E6 - DroughtClip Fig3
Figure 1. Mean squared errors (MSE) for gamma distributed values resulting from different bias correction methods. The MSE of the uncorrected values of 23.5 is efficiently improved with each method. The method used by the central prediction system of MiKlip (“add”; bias correction of the mean) is outperformed from nearly all other methods. However, from “norm” to “dist” similar MSE are reached although the methods strongly differ in their ability to correct higher order moments.
Click to enlarge.

Tasks and deliverables

Different known bias correction methods are compared and an implementation is planned for the central prediction and evaluation system. The utilised approaches include parametric and non-parametric techniques, in particular methods based on distribution functions and the quantile-relations between observed and predicted variables. Further, the development of an improved bias correction is planned, which additionally addresses the problem of climate model drift and reduces time dependent systematic errors. The evaluation of bias correction methods is accompanied by an analysis of prediction scores. They determine efficiently their value from an end-user perspective. The main variable of interest are drought events and how and to which extent bias correction methods help to improve their prediction.

Figure 2. Ranked probability skill score (RPSS) for 3 categories (no precipitation, below and above the median). Positive values show improvement against a reference prediction. Bias correction with “add” does not outperform the climatological prediction (“clim”). The methods “lin” (linear transfer function based on quantile-relation of the observed and predicted value) and “dist” (transfer function based on estimated distribution functions) reach positive RPSS against climatology and even with similar MSE (Figure 1) their RPSS deviate substantial from each other.
E6 - DroughtClip Fig3
Figure 3. Continuous ranked probability skill score (CRPSS), which compares the forecast and observed distribution function. Positive values show improvement against a reference prediction. Better calibrated ensembles can be achieved with improved bias correction methods.

Progress so far

A set of bias correction methods have been selected from published literature. They encompass a wide range of different strategies to reduce systematic departures. On basis of a simulation study the following main results are established:

  • Differences in the performance of bias correction methods for skewed variables can not be addressed with distance measures, like the mean squared error or the mean absolute error. As long as the mean and the variance are successfully corrected, similar errors are reached even with large distributional deviations in the corrected quantities (Figure 1). Probabilistic measures, however, are able to uncover the performance differences of bias correction methods (Figure 2 and 3)
  • Bias correction methods accounting for higher order moments improve probabilistic measures: the consistency of the ensemble members, the accurateness of probability forecasts for multiple categories (Figure 2) and they reduce the difference between the forecast and observed distribution function (Figure 3). In summary, bias correction methods not only reduce systematic errors, they additionally result in better calibrated ensembles

Contact

Max-Plank-Insitut für Meteorologie
Wolfgang Müller

Frank Sienz
frank.sienz(at)nospammpimet.mpg.de

Holger Pohlmann

A twentieth-century reanalysis forced ocean model to reconstruct the North Atlantic climate variation during the 1920s

2015 - Clim. Dyn., Vol. 44 (7), pp. 1935-1955

Müller, W. A. | D. Matei, M. Bersch, J. H. Jungclaus, H. Haak, K. Lohmann, G. P. Compo, P. D. Sardeshmukh, and J. Marotzke

Ensemble size impact on the decadal predictive skill assessment

2015 - Meteorologische Zeitschrift Vol. 25 No. 6, pp. 645-655

Sienz, F. | H. Pohlmann, and W.A. Müller

Decadal climate predictions for the period 1901–2010 with a coupled climate model

2014 - Geophys. Res. Lett., Vol. 41 (6), pp. 2100–2107

Müller, W. A. | H. Pohlmann, F. Sienz, and D. Smith