RainfallNet: A Dual-Source of Spatial-Channel Attention Fusion Network for Precipitation Nowcasting

Abstract

Abstract Accurate rainfall prediction is conductive to human life and disaster prevention. Meanwhile, deep learning methods are confirmed to be helpful to improve the accuracy of weather prediction. A novel data-driven neural network is proposed in this work referred as RainfallNet which introduces fusion module based on both radar echo observations and numerical weather prediction (NWP) data. The architecture of the network includes three elements: (1) dual encoders to extract the spatio-temporal feature of the radar echo images and NWP data respectively, (2) parallel attention mechanism combining channel attention and spatial attention to reveal the contribution of each data source and (3) combined loss function joining structural similarity loss, mean square error and mean absolute error with different weight for each rainfall level to further increase the meteorologically assessment metrics. The experiments on South China dataset demonstrate the effectiveness of our model, achieving superior performance on meteorologically assessment metrics over most existing algorithms.