Statistical Downscaling of Temperature Distributions from the Synoptic Scale to the Mesoscale Using Deep Convolutional Neural Networks

Deep learning, particularly convolutional neural networks for image recognition, has been recently used in meteorology. One of the promising applications is developing a statistical surrogate model that converts the output images of low-resolution dynamic models to high-resolution images. Our study exhibits a preliminary experiment that evaluates the performance of a model that downscales synoptic temperature fields to mesoscale temperature fields every 6 hours. The deep learning model was trained with operational 22-km gridded global analysis surface winds and temperatures as the input, operational 5-km gridded regional analysis surface temperatures as the desired output, and a target domain covering central Japan. The results confirm that our deep convolutional neural network (DCNN) is capable of estimating the locations of coastlines and mountain ridges in great detail, which are not retained in the inputs, and providing high-resolution surface temperature distributions. For instance, while the average root-mean-square error (RMSE) is 2.7 K between the global and regional analyses at altitudes greater than 1000 m, the RMSE is reduced to 1.0 K, and the correlation coefficient is improved from 0.6 to 0.9 by the surrogate model. Although this study evaluates a surrogate model only for surface temperature, it probably can be improved by augmenting the downscaling variables and vertical profiles. Surrogate models of DCNNs require only a small amount of computational power once their training is finished. Therefore, if the surrogate models are implemented at short time intervals, they will provide high-resolution weather forecast guidance or environment emergency alerts at low cost.