Interpolation of Annual Maximum rainfall probability distribution using Graph Neural Networks

The modelling of high resolution maximum rainfall depth distribution plays a key role in preventing natural disasters, such as floods. The rain gauges in meteorological stations distributed over the territory already offer a good base to estimate the maximum rainfall depth distribution. However, the resolution at which the distribution is estimated can be increased through interpolation. In this thesis, the problem of maximum rainfall depth distribution interpolation is addressed by using a graph neural network-based approach. Some variations of the proposed algorithm are compared and evaluated on the northern Italy dataset. In particular, the use of a combination of a simple ensemble and temporal ensemble proved to be particularly effective. Finally, the proposed methods show improved results when compared with the classical methods: ordinary kriging, universal kriging and Gaussian process.