AbstractAs modern numeric weather prediction models move towards higher resolution
predictions, the implementation of more unconventional temperature measurements, such as crowd sourced measurements, become an interesting avenue to explore. In this thesis, we attempted to construct machine learning models with the data from the VLINDER citizen science project and the data from the Royal Meteorological Institute of Belgium’s synoptic network of automatic stations. With these machine learning models aimed be able to predict the temperature at a location based on the land cover and meteorological information from the synoptic RMI stations. Several neural networks with different architectures and several non-neural network models, such as a linear regression model, a support vector regression model, a K-nearest neighbour regression model and a random forest regression model were compared against each other on different training sets. It was found that the random forest model was performing almost consistently the best. We also implemented a concept from game theory, called Shapley values, for a post-hoc method to understand how features interact with and are ranked in a model. We were only able to calculate these values for neural network models with decent results.
|Date of Award||Sep 2022|
|Supervisor||Steven Caluwaerts (Promotor) & Lesley De Cruz (Co-promotor)|