Assessing and improving the transferability of machine learning models is an important task for the large-scale mapping of environmental observations. In the context of spatial modelling and remote sensing, the transferability of a model relates to its ability to validly predict in geographic regions where training data is not available. These regions can be identified using the Area of Applicability method currently published and developed in the Uebersat Project (Meyer et al. 2021).
Understanding the influence of the spatial distribution of training data and predictor interactions on the models transferability is crucial for the interpretation of predictions. In Barbiero et al. (2020) the models transferability and generalization from different datasets was tested using a convex hull approach and related to structural characteristics of the dataset (e.g. correlations between predictors).
In this context, we offer a Master Thesis dealing with the following topics:
- Reproduce ideas of Barbiero et al. (2020) and assess model transferability with the Area of Applicability
- Complement the concept by using spatial data and refer model transferability to spatially intrinsic aspects of the data
The Thesis will be part of the Ubersat Project and conducted in a collaborative team from ILOEK and IFGI.
Barbiero, Pietro, Giovanni Squillero, and Alberto Tonda. 2020. “Modeling Generalization in Machine Learning: A Methodological and Computational Study.” arXiv:2006.15680 [Cs, Stat], June. https://arxiv.org/abs/2006.15680.
Meyer, Hanna and Edzer Pebesma. 2021. “Predicting into Unknown Space? Estimating the Area of Applicability of Spatial Prediction Models.” Methods in Ecology and Evolution, June, 2041–210X.13650. https://doi.org/10.1111/2041-210X.13650.
Contact: Marvin Ludwig marvin.ludwig [at] wwu.de