Search for a command to run...
Abstract Precipitation nowcasting, the forecasting of short-term localized precipitation, is a crucial component of weather forecasting, particularly given the increasing frequency of extreme weather events due to climate change. Building on the critical role of precipitation nowcasting, this study focuses on enhancing deep learning models by optimizing hyperparameters on well-established models such as ConvLSTM and SmaAt-UNet rather than proposing new architectures. Our dataset consists of 10 years of precipitation maps at 1x1 km 2 , 5 min resolution from Météo France’s Mosaic product covering the Paris region, providing a robust testbed for training and evaluating models. To better characterize the performance of models over different type of rain event, clustering was carried out in order to partition the events into four different groups ranging from light rain to intense rain. Four main changes were explored to improve the training process. We show that increasing the temporal context up to 40 minutes significantly improves predictive accuracy for high-intensity rainfall events. Second, the introduction of wider spatial observation windows mitigates edge effects during training, enhancing spatial consistency and intensity estimation. Third, extending the output horizon to 90 minutes to predict precipitation at 60 minutes offered substantial gains for recurrent models like ConvLSTM, highlighting their capacity to better capture long-term dynamics. Last, the influence of convolution kernel size was studied to maximize the performance of the models. These changes collectively lead to significant forecast improvements especially for extreme precipitation events with metrics values increasing compared to baseline configurations. We also show that the proposed model performs better when evaluated at coarser scales, relevant for hydrological applications.