Weather Radar Precipitation Nowcasting With Transformer-Based Spatiotemporal Predictive Learning

By Alejandro Salgueiro Dorado

Abstract
In a changing climate, rapid access to precise short-term precipitation forecasts is vital for pre-emptive actions against threats to critical services and human well-being. Traditional physics-based forecasting methods excel in daily predictions but struggle with sudden extreme events due to computational constraints. To address this, deep learning spatiotemporal techniques have emerged as a faster method, notably leveraging novel architectures like transformers. Yet, the application of transformer-based spatiotemporal models to the task of precipitation nowcasting remains largely unexplored. This study seeks to address this gap by evaluating the performance of two distinct transformer-based spatiotemporal models within the framework of precipitation nowcasting. Two transformer-based architectures (SwinLSTM (Tang et al., 2023) & SimVP Swin (Tan, Gao, et al., 2023)) were compared against two convolutional Long Short-Term Memory (LSTM) based architectures (ConvLSTM (Shi et al., 2015) & PredRNNv2 (Wang et al., 2022)). Each architecture was trained on three different configurations with variable complexities and all the models were compared based on their predictive accuracy, nowcasting capability, and computational efficiency. Furthermore, three Mean Squared Error (MSE) derived loss functions were also compared on the SimVP Swin model and showed notable enhancements over the MSE loss function in the model’s nowcasting ability. The findings of this study indicate that all models had their strength, but under the specific training conditions of our experiments, the transformer-based models did not surpass PredRNNv2 accuracy and nowcasting capabilities. However SwinLSTM demonstrated the best potential for enhancement if it received additional training time and incorporated additional mechanisms from PredRNNv2. This research illuminates both the strengths and limitations of the different architectures and loss functions in terms of predictive accuracy and computational efficiency for different precipitation intensities. Moreover, the study also enlightens different directions for combined integration of the presented approaches to foster the development of more robust and efficient nowcasting models. Additionally, the establishment of a Doppler radar pre-processing pipeline can help streamlining the datasets creation to hopefully facilitate future research in this domain.