Towards a better understanding of our Earth with explainable machine learning

Towards a better understanding of our Earth with explainable machine learning
Ribana Roscher,  Junior Professor of Remote Sensing, University of Bonn

LT1 Exchange Building, Monday 11^th July, 3-4pm.

Machine learning approaches, especially deep neural networks, are showing tremendous success in finding patterns and relationships in large data sets for predictions and classifications that are usually too complex to be directly captured by humans. In many cases, machine learning is used to align the learned models as well as possible with our existing knowledge. Also in the field of explainable machine learning, which analyzes the decision process of machine learning methods in more detail, the focus is mostly on matching the models with prior knowledge but less on gaining new scientific knowledge. To enable the derivation of new knowledge, common interpretation methods such as saliency maps are of limited suitability, as they derive interpretations from individual input data to explain the model decision for these specific samples. This talk presents diverse environmental and agricultural sciences applications in which explainable machine learning is used. In particular, a novel approach to the study of wilderness will be addressed, which can be used to derive new knowledge about this land cover class.

http://rs.ipb.uni-bonn.de/people/prof-dr-ing-ribana-roscher/