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.

This additional thesis project is performed as preliminary research for a bigger project that they are going to start at Lund University, to investigate whether the use of a different interpolation methods, to link the precipitation data to the sub-basins centers of the HYPE model, lead to improved model performance. In this report, previously performed research is summarised and the limitations in researching this question with the HYPE model are described. A start is made with investigating this question, by answering the question whether different interpolation methods result in a different discharge when it is assumed that all fallen precipitation ends up as discharge. This is investigated for the PO basin in Italy with 4 interpolation methods: NN, IDW, BIL and OK. The effects of the interpolation methods on the computed discharge time series are analysed with the use of the correlation, RE, NSE and KGE. It is shown that for the PO basin, with a 1000 km2 average sub-basin size and a gridded data set with a resolution of 50 km, the interpolation methods do not produce differences for which you would expect that it could lead to model improvement. However, based on findings from previous research, a next step is proposed, in which we can investigate if we do observe a difference at a different sub-basin scale or data resolution.