Nowcasting high-intensity precipitation is crucial for emergency services and municipalities when making weather-dependent decisions. This research implements and trains a deep generative model for nowcasting using a cleaned precipitation radar composite dataset spanning 15 years, with a 5-minute temporal and 1 km spatial resolution.

We propose and apply a method for speckle-like clutter removal to enhance data quality, particularly for high-intensity precipitation rates. The deep generative model is trained with two data sampling strategies to balance the data and improve the accuracy of high-intensity precipitation forecasts. Model performance is evaluated using several standard metrics, and we propose an adaptation to one metric to quantify a score to peak anticipation time in precipitation nowcasting. We also compare our model's performance with a state-of-the-art deterministic Lagrangian extrapolation-based nowcasting system.

Our results show that the proposed data quality improvement method effectively removes certain errors from historical radar data. Although the deep generative model currently scores low on the standard metrics, the model trained with a focus on high-intensity precipitation shows an improved score to peak anticipation time. Both deep generative models exhibit less blurring compared to the state-of-the-art model and, in some cases, perform similarly or outperform it.

Accurate short term rain predictions are important for flood early warning systems, emergency services, energy management and other services that that make weather dependent decisions. Recently introduced machine learning models suffer from blurry and unrealistic predictions at longer lead times, causing poor performance on the rarer heavy rainfall events. The objective of this research was to explore how the loss function in a recurrent, convolutional neural network (TrajGRU, X. Shi et al. (2017)) can be modified, to get sharper and more realistic predictions, without worsening its performance. Six perceptive loss functions, which should better represent how an image is perceived, are thoroughly analysed to understand the reasons behind their functioning in the context of precipitation nowcasting. It was shown that these perceptive losses can lead to an improvement in sharpness for the first lead time, but that the blurriness for longer lead times arises due to the imbalance in the dataset and the uncertainty of the model with respect to the location of the rain. This thesis gets concluded with recommendations on how to deal with these two problems.