Deep Learning Model based on Multi-scale Feature Fusion for Precipitation Nowcasting

Abstract. Accurate forecast of heavy precipitation remains a challenging task in most deep learning (DL)-based models. This study proposes a novel DL architecture named 'Multi-scale Feature Fusion' (MFF) for precipitation nowcasting for a lead time of up to 3 h. The basic idea is to apply convolution kernels with various sizes to achieve multi-scale receptive fields and then capture the movement features of the precipitation system (e.g. shape, movement direction, and moving speed). Meanwhile, the architecture implants the mechanism of discrete probability to reduce uncertainties and forecast errors, so that heavy precipitations can be produced even at longer lead time. The model uses four year’s radar echo data from 2018 to 2021 for model training and one year’s data of 2022 for model testing. The model is compared with three existing extrapolative models: time series residual convolution (TSRC), optical flow (OF), and UNet. Results show that MFF obtains relatively superior forecast skills with a high probability of detection (POD), low false alarm rate (FAR), small mean absolute error (MAE), and high structural similarity index (SSIM). The most commendable result is that MFF can predict high-intensity precipitation fields at 3 h lead time while the other three models can not. Additionally, it can be found from the results of radially averaged power spectral (RAPS) that MFF shows improvement in the smoothing effect of the forecast field. Future works will pay more attention to multi-source meteorological variables, the structural adjustments of the network, and the combinations with numerical models to further improve the forecast skills of heavy precipitations at longer lead times.