On 8 September at 14:15, Marta Jemeljanova will defend her doctoral thesis, “Spatial machine learning approaches for environmental prediction”, to obtain the degree of Doctor philosophiae in Geography.
Supervisors:
Prof. Evelyn Uuemaa, University of Tartu
Associate Prof. Alexander Kmoch, University of Tartu
Opponent:
Prof. Alexander Brenning, Friedrich Schiller University Jena, Germany
Summary
Machine learning is increasingly used to spatially predict environmental phenomena, especially when field measurements are limited. However, there is a mismatch between how these models work and how environmental data behave: machine learning typically assumes that data points are independent, while in reality, nearby locations tend to be similar. This becomes critical when machine learning is used to support environmental decision-making, as it can lead to overly confident or misleading results. This PhD thesis explores how machine learning can be improved for analysing spatial environmental data, with a particular focus on water quality monitoring. It reviews the main methods for handling spatial dependence and examines broader challenges in applying machine learning to environmental problems. The results show that there is no one-size-fits-all solution: modelling workflows must take into account the spatial distribution of measurements and how well they actually represent the target phenomenon. To produce reliable results, it is important not only to rely on single accuracy metrics but also to validate the spatial patterns of the predictions. Although spatial machine learning has great potential for predicting environmental phenomena, trustworthy results depend on whether the models capture real patterns in nature rather than only patterns in the data.
