November 9-12, 2026 in Leipzig
ELLIS Unit Jena members at the Helmholtz Imaging Conference 2026!
Four days. One community. Endless possibilities in imaging.
From November 9โ12, Helmholtz Imaging Week 2026 brought together the Helmholtz imaging community in Leipzig for an inspiring week of scientific exchange, innovation, and interdisciplinary collaboration.ย At the heart of the week was the Helmholtz Imaging Conference 2026, providing a platform for researchers from diverse scientific fields to present their work, exchange ideas, and explore new perspectives on imaging and its applications. Complementing the conference, the Imaging Deep Dive Day offered interactive, participant-led workshops, creating opportunities to explore selected topics in greater depth and foster direct scientific exchange.
We were delighted to see ELLIS Fellow Prof. Dr. Joachim Denzler (ELLIS Unit Jena Director, Friedrich Schiller University Jena) and ELLIS Fellow Prof. Dr. Miguel D. Mahecha (Associated Member ELLIS Unit Jena, Leipzig University) contribute to this year’s Helmholtz Imaging Conference 2026 in Leipzig with two inspiring keynote talks.
๐๐ผ๐ฎ๐ฐ๐ต๐ถ๐บ ๐๐ฒ๐ป๐๐น๐ฒ๐ฟ โ โ๐๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ ๐๐ป๐๐ฒ๐ด๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ถ๐ป ๐ ๐ฎ๐ฐ๐ต๐ถ๐ป๐ฒ ๐๐ฒ๐ฎ๐ฟ๐ป๐ถ๐ป๐ด: ๐๐ฟ๐ผ๐บ ๐๐ผ๐บ๐ฎ๐ถ๐ป ๐ฎ๐ป๐ฑ ๐ฃ๐ต๐๐๐ถ๐ฐ๐ฎ๐น ๐ฃ๐ฟ๐ถ๐ผ๐ฟ๐ ๐๐ผ ๐๐ฎ๐๐๐ฎ๐น ๐๐
๐ฝ๐น๐ฎ๐ป๐ฎ๐๐ถ๐ผ๐ป๐โ
Machine learning can identify patterns in complex scientific data, but its predictions may depend on shortcuts, biases, or correlations that fail outside the training setting. This talk explores how established knowledge can guide learning and make model behavior easier to understand. First, it shows how expert class hierarchies can shape image embeddings so that visual retrieval reflects semantic relationships. It then examines physics-informed methods for recovering Raman spectra from coherent anti-Stokes Raman scattering measurements, using known spectral relationships to reduce dependence on labeled data. Finally, the presentation introduces causal reasoning as a way to examine what models rely on, how they respond to interventions, and what changes might alter an individual prediction. Together, these cases show how domain, physical, and causal knowledge can support more robust and interpretable AI for scientific applications.
๐ ๐ถ๐ด๐๐ฒ๐น ๐ ๐ฎ๐ต๐ฒ๐ฐ๐ต๐ฎ โ โ๐๐ฟ๐ผ๐บ ๐ฃ๐ถ๐
๐ฒ๐น๐ ๐๐ผ ๐ฃ๐น๐ฎ๐ป๐ฒ๐๐ฎ๐ฟ๐ ๐๐๐ป๐ฎ๐บ๐ถ๐ฐ๐: ๐ง๐๐ฟ๐ป๐ถ๐ป๐ด ๐๐ฎ๐ฟ๐๐ต ๐๐บ๐ฎ๐ด๐ฒ๐ ๐ถ๐ป๐๐ผ ๐๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ ๐ฎ๐ฏ๐ผ๐๐ ๐ผ๐๐ฟ ๐๐ต๐ฎ๐ป๐ด๐ถ๐ป๐ด ๐๐ถ๐ผ๐๐ฝ๐ต๐ฒ๐ฟ๐ฒโ
The Earth is experiencing extraordinary transformations at an unprecedented pace. We can observe many of these changes from above, using drones and satellites. The resulting images of the Earthโs surface have become a primary tool for monitoring change, allowing us to observe the terrestrial biosphere at extraordinary spatial and temporal detail. However, the challenge is not only to detect objects and describe states, but to understand the dynamic processes they represent. In our research, we move from detecting, for example, tree mortality in centimetre-scale drone imagery to analysing forest health dynamics across space and time using dense satellite time series. This requires moving beyond individual images towards large spatio-temporal and multivariate Earth System Data Cubes. Deep learning and interactive visualisation can reveal structures and trajectories within these high-dimensional observations and increasingly allow us to infer changes in ecosystem state at large scales. I will illustrate this progression from objects to dynamics and finally to planetary-scale patterns with the global green wave, where decades of satellite observations can be condensed into an interpretable trajectory of the changing seasonal dynamics of the terrestrial biosphere.