Master Thesis Defense by Jonathan Ortved Melcher

Title: Using Novel Explainable Machine Learning Methods to Study Mean Precipitation in Denmark

Abstract:
This thesis explores the use of eXplainable Machine Learning (x-ML) to expand the knowledge of connections between precipitation in Denmark and meteorological phenomena in the North Atlantic. While current research on the North Atlantic Oscillation (NAO) mainly employs linear correlations and empirical orthogonal functions, this thesis attempts to capture neglected nonlinear connections.

 Several Neural Networks (NN) are trained to model the mean monthly precipitation over Denmark using surface temperature, mean sea level pressure, the 500 hPa geopotential height and the length of its gradient over the North Atlantic. The data comes from 17 EC-Earth ensemble members within the CMIP6 historical framework (1850-2014). To ensure EC-Earth's capability of capturing the established NAO signal, these are reproduced through traditional statistical methods. The x-ML technique, Layerwise Relevance Propagation (LRP), was applied to identify meteorological variables and locations relevant for the modeling of precipitation.

Nineteen different models and data setups were tested, changing input variables, preprocessing techniques, NN architecture, seasonal constraints, input regions and moving the target area from Denmark to other locations. During the development of this thesis, a switch was made from a classification to a regression task as this better reflected the precipitation patterns. Despite extensive experimentation with feed forward neural networks, supplemented by convolutional neural networks and k-means clustering of relevance patterns, a consistent finding emerged: local mean sea level pressure dominates the relevance for precipitation modeling. This shows pitfalls in using LRP for exploring novel nonlinear patterns in data sets with strong correlations.

This thesis identifies challenges in applying LRP to find nonlinear climate patterns and suggests future work using daily time resolution and time-lagged input to break the linear correlation and better capture meteorological phenomena. The thesis reveals an interesting correlation between the length of the gradient of the 500 hPa geopotential height over northern France and precipitation in Denmark that needs further research.

Supervisor: Jens Hesselbjerg Christensen

Censor: Peter L. Langen (Aarhus Universitet)