Statistical Post-Processing for Gridded Temperature Prediction Using Encoder-Decoder-Based Deep Convolutional Neural Networks
This work addresses temperature prediction challenges for meteorological agencies, representing an incremental improvement over existing Kalman filter methods.
The study tackled the problem of correcting gridded temperature predictions from numerical weather models, which struggle with front location errors and extreme temperatures, by proposing an encoder-decoder-based convolutional neural network. The model greatly improved operational guidance and corrected biases, though no specific numerical improvements were provided.
The Japan Meteorological Agency operates gridded temperature guidance to predict two-dimensional snowfall amounts and precipitation types, e.g., rain and snow, because surface temperature is one of the key elements to predict them. Operational temperature guidance is based on the Kalman filter, which uses temperature observation and numerical weather prediction (NWP) outputs only around observation sites. Correcting a temperature field when NWP models incorrectly predict a front's location or when observed temperatures are extremely cold or hot has been challenging. In this study, an encoder-decoder-based convolutional neural network has been proposed to predict gridded temperatures at the surface around the Kanto region in Japan. Verification results showed that the proposed model greatly improves the operational guidance and can correct NWP model biases, such as a positional error of fronts and extreme temperatures.