A bias-correction and downscaling technique for operational extended range forecasts based on self organizing map

If a coarse resolution dynamical model can well capture the large-scale patterns even if it has bias in smaller scales, the spatial information in smaller domains may also be retrievable. Based on this hypothesis a method has been proposed to downscale the dynamical model forecasts of monsoon intraseasonal oscillations in the extended range, and thus reduce the forecast spatial biases in smaller spatial scales. A hybrid of clustering and analog technique, used in a self organizing map (SOM)-based algorithm, is applied to correct the bias in the model predicted rainfall. The novelty of this method is that the bias correction and downscaling could be done at any resolution in which observation/reanalysis data is available and is independent of the model resolution in which forecast is generated. A set of composite pattern of rainfall is identified by clustering the high resolution observed rainfall using SOM. These set of composite patterns for the clustered days in each cluster centers or nodes are saved and the model forecasts for any day are compared with these patterns. The closest historical pattern is identified by calculating the minimum Euclidean distance between the model rainfall forecast and the observed clustered pattern and is termed as the bias corrected SOM-based post-processed forecast. The bias-corrected and the SOM-based reconstructed forecasts are shown to improve the annual cycle and the skill of deterministic as well as probabilistic forecasts. Usage of the high resolution observational data improves the spatial pattern for smaller domain as seen from a case study for the Mahanadi basin flood during September 2011. Thus, downscaling and bias correction are both achieved by this technique.