We have implemented a CNN (convolutional neural network) to downscale coarse-resolution (2°) upper atmosphere variables to high-resolution surface variables (1/4°) over the central United States. It takes 8 inputs (temperature, humidity, winds, and geopotential height at different pressure levels) and generates 7 outputs (precipitation, minimum and maximum temperature, incoming solar radiation, and surface winds). This set of input and outputs matches those used in our work on differential credibility analysis of downscaling methodologies, and allows direct comparison with both statistical and dynamical models.

The network is trained on inputs from ERA-Interim for 1980-2010, targeting gridMET observations. It is then tested on data from 2011-2016. After training, we apply the model to inputs from the CMIP5 MPI-ESM-LR simulation. We evaluate outputs for the month of May.

We found that using the Structural Similarity Image Metric (SSIM) instead of RMSE (root mean-square error) for the loss function dramatically reduced visual artifacts in the output. The addition of the cosine of the day of the year as an input also improved results, as did adding a missing-data mask as an output channel.

Based on error in out-of-sample predictions and visual inspection, initial results are promising. To more meaningfully evaluate the CNN's credibility, we use partial dependence plots. We split the daily data into three classes: dry, wet, and moist days (precipitation below trace and above and below the 70th percentile, respectively) based on the average downscaled rainfall over a region in the Southern Great Plains. We then perform process analysis on the average upper conditions for those classes to determine whether they are capturing the important physical processes known to control precipitation in this region.