Statistical Post-processing for Gridded Temperature Forecasts Using Encoder-Decoder Based Deep Convolutional Neural Networks

Japan Meteorological Agency (JMA) has been operating gridded temperature guidance for predicting snow amount and precipitation type because those elements are susceptible to a temperature at the surface. The operational temperature guidance is based on the Kalman filter technique and uses temperature observation and NWP outputs only at observation sites; it has been difficult to correct a temperature field when NWP models did not predict the location of a front correctly or when the observed temperature was extremely cold or hot. In the present paper, encoder-decoder-based convolutional neural networks (CNNs) were employed to predict gridded temperatures at the surface around the Kanto district. The verification results showed that the proposed method improves operational guidance significantly and can correct NWP model biases, including a positional error of fronts and extreme temperatures.