Radar rainfall nowcasting stands for the prediction of rainfall amounts and intensities over the next 6 hours by means of statistical extrapolation of radar measurements. It is the principal ingredient for modern flood forecasting and early warning systems. Radar forecasts are ge
...

Radar rainfall nowcasting stands for the prediction of rainfall amounts and intensities over the next 6 hours by means of statistical extrapolation of radar measurements. It is the principal ingredient for modern flood forecasting and early warning systems. Radar forecasts are generated by identifying and tracking rainfall cells in radar images and extrapolating them along the main direction of motion based on the assumption of Lagrangian persistence. The latter means that the rainfall cells do not physically evolve over time. Therefore the key to improving predictions is to anticipate the growth and decay of rainfall cells during the next few hours. Studies have shown that estimating growth and decay based on past radar images is nearly impossible. Therefore, in this thesis, a new approach for predicting growth and decay based on physical guidance from a numerical weather prediction model is investigated. The idea is that numerical weather prediction models are better at anticipating changes in the atmosphere than radar and thus, should also contain information about the growth and decay of rain cells. To test this hypothesis, two simple machine learning models have been trained to learn the relationship between the observed growth and decay in radar images and the output parameters from the Dutch mesoscale numerical weather prediction model HARMONIE. The models were trained on data in the summer of 2019 around the city of Rotterdam for two area sizes, $20\times 20$ km$^2$ and $60\times 60$ km$^2$. The first model is static and assumes that all predictions from HARMONIE are correct. Unfortunately, this turns out to be too optimistic as HARMONIE often places the rain cells at the wrong locations with the wrong intensities and trends. As a result, no meaningful relationship between HARMONIE outputs and growth/decay in radar could be learned. The dynamic model overcomes this issue with the help of an additional classifier. The classifier predicts whether the information from HARMONIE can be trusted or not. Then, a regression model predicts the magnitude of growth and decay (only for the trusted cases). Out of 308 analysed cases, the classifier labelled 237 HARMONIE predictions as being untrustworthy, thereby removing a lot of bad cases. However, 98 of these were false negatives, meaning that they could have been used to predict growth and decay. On the other hand, only 71 cases were labelled as containing useful information, from which 27 cases were false positives. Despite these errors, the dynamic model was able to improve the root mean square error on the predicted growth/decay by $27$\% for these 71 cases. Also, the correlation between the predicted and actual growth/decay increased from near zero to $0.488$. This is encouraging and clearly highlights the potential of this approach. Still, some important challenges remain. In particular, results show that the dynamic model often underestimates the magnitude of growth and decay. Also, when the entire validation data set is considered, including all unusable cases, the improvement with respect to the static model is only $6$\%. This is mainly caused by the poor performance of HARMONIE and the large number of unusable cases. However, performance could be improved further by building a better classifier capable of identifying all good cases with a low number of false negatives.