Development of latent resampling downscaling and its application to model bias and climate change projection

We have developed a cost-effective downscaling technique called latent resampling downscaling (LRDS) to address model uncertainty in regional climate change assessments. In this study, the LRDS was used to generate a surrogate dynamical downscaling (DDS) dataset for a coupled model intercomparison project phase 5 global climate model. This was achieved by sampling from a large ensemble of DDS datasets from d4PDF project. The sampling process was guided by the probability density functions of the global model’s weather patterns, which were classified using a self-organizing map algorithm. We applied LRDS to investigate summertime precipitation over Kyushu Island, Japan. The present-climate simulation revealed considerable inter-model variety in reproducing sea level pressure patterns around Japan in boreal summer. A storyline approach was employed to characterize three distinct behaviors of LRDS in simulating climatological and extreme rainfall over Kyushu under the present climate. Using LRDS, we also evaluated the projected response to climate change. Two contrasting storylines were identified: one showing an increase in rainfall over western Kyushu, and the other indicating increased rainfall over eastern Kyushu.