MLTRANS: Transferability of machine learning models in digital soil mapping

MLTRANS: Machine learning models have shown great success in learning complex patterns such as the spatial distribution of soil properties, making predictions about areas not covered. In contrast, the ability to apply what is learned to other domains is poorly developed. To date, the models have very limited applicability to domains outside the training learning environment. Similar to empirical regressions, the rule sets, e.g. for decision tree methods like Random Forest, only apply to the range of values covered by training data. Training data as high quality and extensive as possible for each additional domain is needed again. Advances in Deep Learning (DL), e.g. Convolutional Neural Networks, Transfer Learning, and combined approaches in Feature Selection (FS) offer here extended possibilities to constrain dimensionality, especially for smaller datasets, to minimize overfitting to the training data, and to improve transfer to adjacent domains. In this application, we incorporate these developments and attempt to predict soil properties for areas outside the learning environment.

Overview  
Project Transferability of machine learning models in digital soil mapping
Start/End 2022-2026
Funding Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)
Keywords Digital Soil Mapping, Deep Learning, Feature Selection, Extrapolation
Investigators and affiliations Nafiseh Kakhani (Universität Tübingen)
Ruhollah Taghizadeh (Universität Tübingen)
Thomas Scholten (Universität Tübingen)
Contact Nafiseh Kakhani, Ruhollah TaghizadehThomas Scholten