PCEED-Net: A Physics-Constrained Enhanced Encoder–Decoder Network for Precipitation Nowcasting

Precipitation nowcasting provides precipitation intensity and coverage estimates over a short future period. Accurate precipitation nowcasting is essential for agriculture, flood warnings, and transportation management. However, as the forecasting time increases, the existing deep learning-based nowcasting methods tend to exhibit obvious forecasting errors. To address this limitation, a physics-constrained enhanced encoder-decoder network called PCEED-Net is proposed in this paper. PCEED-Net adopts an enhanced encoder-decoder architecture, which incorporates the normalized spatiotemporal convolutional memory networks (NSCM-Nets) and the spatial-channel dual attention (SCDA) modules for precipitation nowcasting. These NSCM-Nets enhance the multiscale temporal precipitation and cloud-top temperature features extracted from the main encoder. The SCDA module further guides the network to focus on important features. Additionally, to better capture the spatiotemporal evolution trends of precipitation and constrain the learning process of the model, a new physics-constrained loss function is designed. Focusing on the coastal regions of China and Southeast Asia, the PCEED-Net uses global precipitation measurement (GPM) precipitation data and the Himawari satellite data to perform hourly precipitation forecasting. The proposed model outperforms other existing models in terms of precipitation nowcasting tasks, with average probabilities of detection (PODs) of 0.83, 0.45, and 0.24 and average critical success indices (CSIs) of 0.43, 0.34, and 0.19 at thresholds of 0.1 mm/h, 1.6 mm/h, and 8.0 mm/h, respectively. The results highlight the potential of the proposed method for conducting high-precision precipitation forecasting in the observation regions.