Title: Enhancing Precipitation Nowcasting via Multi-Quantile Regression Beyond MSE

Abstract: Deep-learning approaches for precipitation nowcasting typically rely on pointwise loss functions, such as mean absolute error (MAE) or mean squared error (MSE). However, this optimization strategy often results in forecasts that are excessively smoothed and fail to adequately capture intense rainfall events. This research explores whether the predictive accuracy of a well-established deterministic nowcasting framework can be enhanced by shifting the training paradigm to a multi-quantile regression task. Utilizing the SmaAt-UNet architecture as the foundational model, we evaluate the performance of multi-quantile pinball-loss training against traditional MSE and MAE methods using radar data from the Netherlands. Our analysis demonstrates that multi-quantile training not only refines the central deterministic forecast—reducing test-set MSE by 8.6% relative to an MSE-trained baseline—but also generates upper-quantile predictions that are valuable for risk-aware forecasting of heavy precipitation. These outcomes indicate that quantile regression serves as a straightforward substitute for conventional pointwise losses, eliminating the need for architectural modifications or generative sampling techniques. The code for our models and the associated training configuration can be accessed at \href{https://github.com/gijsvn/Multi-Quantile-Precipitation-Nowcasting}{GitHub}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC