Quantitative Short-Term Precipitation Model Using Multimodal Data Fusion Based on a Cross-Attention Mechanism

Abstract

Short-term precipitation prediction through abundant observation data (ground observation station data, radar data, etc.) is an essential part of the contemporary meteorological prediction system. However, most current studies only use single-modal data, which leads to some problems, such as poor prediction accuracy and little prediction timeliness. This paper proposes a multimodal data fusion precipitation prediction model integrating station data and radar data. Specifically, our model consists of three parts. Firstly, the radar feature encoder comprises a shallow convolution neural network and a stacked convolutional long short term memory network (ConvLSTM), which is used to extract the spatio-temporal features of radar-echo data. The weather station data feature encoder is composed of a fully connected network and an LSTM, which is used to extract the sequential features of the weather station data. Then, the cross-modal feature encoder obtains cross-modal features by aligning and exchanging the feature information of the radar data and the weather station data through the cross-attention mechanism. Finally, the decoder outputs the quantitative short-term precipitation prediction value. Our model can integrate station and radar data characteristics and improve prediction accuracy and timeliness, and can flexibly add other modal features. We have verified our model on four short-term and impending rainfall datasets in South Eastern China, achieving the best performance among the algorithms.