Abstract
The downscaling method for hydrometeorological field is developed, using deep learning. The models are 4 layer feed-forward artificial neural networks which predict RCM's surface 2m temperature and precipitation fields from GCM's field. They are trained by stochastic gradient descent method, using the back propagation method. For initialization, stacked autoencoder was used. The developed models are applied to an area around Japan. The downscaling result represents the spatio-temporal variation of RCM's surface 2m temperature and precipitation well. This implies the effectiveness of our method for the emulation of RCM's dynamical downscaling with low calculation cost.
