Statistical Downscaling

My research involves the use of machine-learning methods (e.g. Support Vector Regression, Bayesian Neural Networks, Probabilistic Graphical Models, CART) to downscale coarse resolution datasets to local scales.

In particular, I'm interested in the study of the often overlooked stationarity assumption, common to all statistical downscaling methods, and its impact on future local projections.

My current research is related to GFDL's "Perfect Model" evaluation framework, where  outputs from a high-resolution GCM are used as "observations", and a coarsened versions of these outputs are used as predictors for the statistical model; thus allowing the downscaled outputs to be evaluated versus the historical and future high-resolution "observations".

For more details about the "Perfect Model" approach visit our group web page: https://www.gfdl.noaa.gov/esd_eval

Also, I just made publicly available the Statistically Downscaled Dataset for the Red River Basin. The dataset contains downscaled historical and future projections at 1/10th of a degree for the Red River of the South Basin (South Central U.S.A). The idea is that practitioners will realize that the local scale climate projections are also dependent on the statistically downscaled method used (i.e. not all downscaling methods are the same). For more information about this novel dataset, please visit the Publications section.