Applications of Self-Organizing Maps to Statistical Downscaling of Major Regional Climate Variables
Yin, C. (2011). Applications of Self-Organizing Maps to Statistical Downscaling of Major Regional Climate Variables (Thesis, Doctor of Philosophy (PhD)). University of Waikato, Hamilton, New Zealand. Retrieved from https://hdl.handle.net/10289/5733
Permanent Research Commons link: https://hdl.handle.net/10289/5733
This research developed a practical methodological framework, which integrated most of the important aspects related to statistical downscaling. The framework showed high skills when applied to downscale daily precipitation, minimum and maximum temperatures over southeast Australia. Within the framework, self-organizing maps (SOM) algorithm was incorporated as the core technique for interpreting the relationship between the predictor and predictand under consideration following the latest advances in synoptic climatology. The SOM classified large-scale predictors into a small number of synoptic patterns on a physically meaningful basis. By mapping the observed local climate variable (predictand) to these patterns, a downscaling model structure, SOM-SD, was constructed based on the NCAR/NCEP reanalysis data. Moreover, for a new atmospheric state, an ensemble of predictand values was generated by a stochastic re-sampling technique inside the SOM-SD. To improve seasonality of downscaled results, a simple seasonal predictand pool (SPP) scheme was introduced, which can acquire similar skills as the traditional solutions of dividing a year into four seasons. The framework identified and applied a broad suite of statistical indices, including mean, variance, cumulative distribution function (CDF), extreme events to assess the performance of the SOM-SD. In addition, some non-parametric methods were also employed to evaluate the uncertainty of the downscaling approach, which improved its robustness in practice. The quality control of the input data consists of another important component of the framework, which assessed GCM predictors from three aspects: (a) replicate reliably synoptic patterns depicted by the reanalysis data; (b) remain relatively stable in the future; and (c) produce similar downscaling skills as the reanalysis data. Finally, the framework provided an equal-distance CDF mapping method to adjust the discrepancies between the downscaled values and the corresponding observations. This method adjusted the downscaled CDF for the projection period on the difference between the CDFs of the downscaled GCM baseline and observed values. Thus the framework combines the advantages of statistical downscaling model and bias correction method. Moreover, the framework puts a strong emphasis on its flexibility, which underpins its application to other regions, as well as to support impact assessment studies.
University of Waikato
All items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.
- Higher Degree Theses