Methodological Focus on Hyperparameters for Different Rain Nowcasting Models

Precipitation nowcasting plays an essential role in operational weather forecasting services. Sudden precipitation events have significant socio-economic impacts, including natural disasters like flash floods. This challenge is becoming increasingly critical as climate change alters weather patterns and the frequency of extreme weather events continues to increase.

Over the last decade, radar observations, offering high temporal and spatial resolution, have facilitated the development of machine learning methods for precipitation nowcasting. Once trained, these methods are well suited to processing large datasets with low latency, especially in a real-time context. Recent advances in the field of nowcasting have focused on optimizing model architectures, improving loss functions for imbalanced data, and integrating multivariate inputs, including radar and satellite observations.

This study explores some critical hyperparameters, such as temporal context length, edge effect during training, influence of the output horizons prediction, and convolution kernel size. To do this, we investigate the performance of several models, including both machine learning approaches from different families, in particular SmaAt-Unet, ConvLSTM , and DGMR (trained on UK rains) , as well as non-machine learning methods such as  STEPS. An eleven years consistent radar precipitation dataset covering the Paris region was set up from Météo-France mosaic. Nine years were used for training machine learning models, and two years were reserved to evaluate the models’ performances. To assess the model in different weather conditions, the data set is divided into four groups with distinct characteristics corresponding to various meteorological phenomena. To ensure consistent evaluation, we evaluated the models on the same two-year test dataset, focusing on three criteria, namely: spatial consistency (Pearson correlation coefficient), location accuracy (CSI), and precipitation intensity (MSE).

Our analysis reveals that machine learning models consistently outperform traditional optical flow methods, with notable variations in performance across timescales and rainfall intensities. We also highlight that performance is nearly identical for all models in the presence of stratiform rain, while there are substantial differences in the convective rain group. Additionally, we show that for deep learning models, considering edge effects during training prevents the propagation of inevitable errors and helps avoid the appearance of ghost rain cells at the edges of the map. Furthermore, we show that the size of the kernels of the first layers plays an important role and must be large enough to allow correlation between distant pixels.

Finally, our study provides guidelines for the development of precipitation nowcasting models.