Data-driven dynamic motion field generation for rainfall nowcasting

Rainfall nowcasting algorithms rely primarily on extrapolation, where recent radar rainfall observations are projected forward in time based on a motion field that is determined with past data. While additional (stochastic) processes may be incorporated, as is for example done in the pySTEPS models, extrapolation remains the fundamental mechanism. Although the motion field estimates are robust, they assume a steady state in the motion field for the future. This assumption can face significant challenges in maintaining accuracy over time, especially during convective weather events characterized by rapid changes in precipitation patterns and their movement.

In this study, we focus on three objectives: 1) identifying the current errors and uncertainties in the steady-state motion field derivation using pySTEPS, 2) the construction of a dynamic motion field derivation approach using a new deep-learning model, MotioNNet, and 3) the development of ensemble motion fields for MotioNNet. MotioNNet is a U-Net based deep-learning architecture, which uses the past radar images (five in this study) in combination with the estimated static motion field from pySTEPS to estimate the deviation from the provided static motion field per grid cell with increasing lead time. For the ensemble generation in MotioNNet, we tested probabilistic techniques such as SpatialDropout and Monte Carlo dropout.

We trained and tested our model on C-band weather radar data from the Royal Netherlands Meteorological Institute (KNMI), using 10,000 rainfall events. These events were selected to include cases with both intense precipitation and significant motion errors. Our results show that the static motion field approach results in average motion field errors of 1 – 3 km h^-1 at the start of the forecast and increases to 4 – 8 km h^-1 (on average, and locally sometimes much higher) at a lead time of 90 minutes. The dynamic motion field estimates of MotioNNet improve the motion prediction accuracy by approximately 13%. The improvement is much higher for structured and stable events (up to 45%), but almost negligible for localized thunderstorm events. The results of the ensemble construction in MotioNNet indicate that MotioNNet is capable of adding perturbations in space where most uncertainty takes place, especially for structured and stable events. This is an advantage compared to the spatially uniform approach of pySTEPS. However, the spread of the ensembles is still underestimated, even more so than with pySTEPS, indicating that the uncertainty in the forecast is not yet well represented.

We conclude that the hybrid MotioNNet approach can substitute and enhance parts of the motion field module in pySTEPS. MotioNNet refines initial motion field estimates, rather than replacing them, which leads to a modular approach that fits well in the overall pySTEPS framework. We expect that the dynamic motion field approach from MotioNNet will aid in further enhancing the predictability of (high-intensity) rainfall events for short lead times, especially for structured events where motion errors currently play a role in the forecast error.