Extreme rainfall brings substantial threats to lives, infrastructure, and the economy in cities. Radar rainfall nowcasting was proven able to provide forecasts up to 2 to 3 hours in advance on a catchment scale. However, an extensive evaluation of nowcasting skills for urban areas has not been performed yet. In this study, we selected 80 extreme events that occurred in 5 main Dutch cities (Amsterdam, The Hague, Groningen, Maastricht, and Eindhoven) from 2008 to 2021. We evaluated the performance of probabilistic nowcasts with 20 ensemble members applying short-term ensemble prediction system (STEPS) from Pysteps for these cities, focusing on analyzing the dependence on rainfall characteristics and city sizes. Nowcasts in Eindhoven (96 km2) and Maastricht (67 km2) had higher errors because the rainfall intensity of their events was higher. Besides, nowcasts at small areas showed higher error, especially when the size was below 100 km2. We found that forecast errors were higher and the forecast was less reliable for the 1-h event durations than for 24-h durations. Despite these differences, skillful lead times measured by Pearson correlation in all the cities were about 20 to 24 minutes for both the 1-hour and 24-hour events. CARROTS (Climatology-based Adjustments for Radar Rainfall in an Operational Setting) adjusted the bias in real-time QPE and QPF, but QPF still reduced with increasing lead time. Also, CARROTS did not adjust the rainfall spatial distribution much, so the skillful lead time did not change much. The skillful lead time in this study was shorter than the counterparts on the catchment scale because small areas are more sensitive to the displacement of forecast rainfall. Still, such lead time is similar to the findings in other research on short convective rainfall over small areas. Future research could try to apply machine learning, 3-dimensional nowcasting, or blending numerical weather prediction in the nowcasting process to better forecast the growth and decay of rainfall at a longer lead time.

This thesis aims to enhance rainfall nowcasting by improving motion field predictions and ensemble generation within PySTEPS using machine learning techniques. Accurate nowcasting is crucial for flood early warning, agriculture, transportation, and public safety. The steady-state assumptions and ensemble generation in PySTEPS, and many other nowcasting methods, face significant challenges in maintaining accuracy over time, especially during convective weather events characterized by rapid changes in precipitation patterns and their movement. The research focuses on three objectives: identifying the current errors and uncertainties in PySTEPS motion field estimations, constructing dynamic motion field predictions with the development of a new deep-learning model, MotioNNet, and developing motion field ensemble generation methods for MotioNNet. We introduce a new dataset containing 10,000 motion field events, consisting of events with high rainfall and motion errors in the NL. MotioNNet, trained/validated/tested on 75%/12.5%/12.5% of the dataset, improves motion prediction accuracy by approximately 13%. It provides accurate temporal adjustments (e.g., advecting the motion), especially in structured and stable events. For ensemble generation, probabilistic techniques such as SpatialDropout and Monte Carlo dropout are employed within MotioNNet. The results indicate better ensemble predictions in structured and stable events, accurately identifying problematic areas and generating realistic members. However, it presents a small spread compared to PySTEPS. Challenges remain in accurately capturing uncertainty during highly dynamic convective events, but new paths are open for future enhancement. The study concludes that integrating machine learning models into traditional nowcasting frameworks can substantially enhance motion field predictions and ensemble generation. Future work should focus on further refining these models or exploring more advanced architectures (e.g., ConvLSTM, diffusion models, or transformers), improving context-dependent perturbations in ensemble generation, and enhancing the references and assessments. These advancements hold potential for improving motion field nowcasting accuracy.