Abstract
In this study, we developed deep learning methods using a convolution neural network (CNN) including shortcut paths (U-Net) to correct numerical weather prediction model outputs. These methods output the distribution of precipitation intensity, and can be used as precipitation guidance in short-term precipitation forecast.
As input data, the precipitation intensity and vertical near-surface wind speed predicted by the CReSiBUC numerical weather prediction model in the Keihanshin region of Japan between August 1 and 31, 2001, were used and splited into training and validation data. We used the observed precipitation based on the Radar/Raingauge-Analyzed Precipitation provided by the Japan Meteorological Agency. The observed data were interpolated into the same meshes using the nearest neighbor method simultaneously with the prediction results.
The two-dimensional rainfall distributions before and after being corrected by the U-Net and the CNN without shortcut paths were compared with the observed data from the Radar Raingauge Analyzed Precipitation. In the case of precipitation with a large spatial scale (each precipitation area extends around several tens km in diameter; e.g. typhoon), both the CNN and U-Net correct the precipitation distributions in the areas where the numerical meteorological model predicted no precipitation. In addition, the U-Net can express the region where the precipitation peaks (where precipitation intensity exceeds 5mm/hr or 10mm/hr), which are important for disaster mitigation. On the other hand, in the case of localized precipitation (where each precipitation area extends several kilometers in diameter), it was found that the displacement of the precipitation position could not be completely corrected both in CNN and U-Net, so further improvement will be required.
