Lightweight residual U-Net model for hourly precipitation nowcasting

Precise and accurate forecasting of rainfall plays a crucial role during or before extreme weather events like thunderstorms, floods, landslides etc. With the progress in machine learning and the abundance of spatiotemporal meteorological data, deep learning-based weather forecasting has garnered significant attention. Building lightweight deep learning models is a recent focus in computer science to reduce training time and computational costs while achieving effective results. In recent years, the U-Net architecture has been widely used for the forecasting of weather variables, including precipitation. This paper proposes a lightweight U-Net model fused with residual blocks which include convolutional layers and separable convolutions for effective precipitation nowcasting. The proposed architecture receives the previous three hours of precipitation data as inputs and outputs the precipitation in the next 1-2 hours horizon. The effect of loss functions and the time horizon on the model predictions have also been studied in the paper. The model has been trained and evaluated for nowcasting over the northeastern region of India during the months of June, July, August, and September using two different hourly precipitation datasets, viz. the ERA5 reanalysis and GSMaP operational with spatial resolution 0.25 ◦ and 0.1 ◦ respecticvely. The devised architecture achieves a best R 2 of 0.718, RMSE of 0.54, MSE of 2.977e-4, MAE of 0.21 milimeters per hour and a NMAE of 0.32% for nowcasting within a 1-hour horizon trained using the MAE loss function with the Era5 reanalysis dataset, demonstrating its effectiveness for precipitation prediction. Evaluation of the model’s performance and comparison with a convolutional U-Net architecture, reveal that the proposed model produces significantly better results with a much lesser number of parameters.