Smart Data & Knowledge Services

In the topic area "Earth and Space Applications" techniques of machine learning are used to deepen  our understanding of proccesses on our planet and i surrounding space environment. In particular, the team focuses on the use of earth observation data in combination with artificial intelligence for different application areas such as agriculture, urban planning, air pollution monitoring and flood mapping.

Multi-modal learning

In earth observation, there is often an immense number of relevant data sources that can contribute to answering a particular research question. Common input data sources are multi-spectral optical satellite images but also radar images, weather data and digital elevation models. We are therefore concerned with the development of machine learning models that can combine different data sources with often different spatial and temporal resolutions. In general there are two approaches to this problem: early fusion and late fusion. If you use early fusion, the data sources are combined with each other at an early stage. The models can learn best from the correlations of the data, but at the same time the temporal and spatial scale of the data must be harmonized. This is not necessary with late fusion, since the data is first processed separately in the model. The different versions are only combined when the final result is calculated.

Transfer learning

While there are a variety of input data sources in Earth observation, annotating data correctly is a challenge that can often only be done manually. For many application areas, there are therefore only small annotated data sets for training machine learning models. We work on the transfer of models from one application area to another or from one region to another. This allows us to minimize the amount of training data in new application areas and regions.

Bridging the gap to applications

We focus on the direct industrial application of our research results. We therefore collaborate closely with domain experts for each of our application areas.