Fachsymposium „Artificial Intelligence for Life“ 2026

Anna Mullaney
University of Applied Sciences Weihenstephan-Triesdorf 
Deep Learning on Blood NMR Spectra for Health Assessment

Anna Mullaney, Raphael Kozlovsky, Christina Krempl, Clemens Thielen, Andreas Krumpel, Tobias Ameismeier and Roland Geyer
Technical University of Munich (TUM); lifespin GmbH, Regensburg; Hochschule Weihenstephan-Triesdorf (HSWT), Germany

The molecular composition of human blood can be characterized in detail using proton (¹H) nuclear magnetic resonance (NMR) spectroscopy, which holds considerable potential for personalized health assessment. Conventional analysis pipelines reduce spectral data to a fixed set of predefined biomarkers, discarding much of the underlying biochemical information. Deep learning offers an alternative by extracting meaningful patterns from the full spectrum, without relying on predefined features. However, its application to blood NMR has been hindered by the scarcity of annotated training data. Self-supervised approaches address this challenge by learning from the structure of the data itself.

In this work, we investigate self-supervised deep learning as a framework for learning directly from blood NMR spectra. Leveraging more than 70,000 serum spectra spanning several hundred disease contexts, this approach integrates experimental and realistically simulated spectra to enable large-scale pretraining without extensive manual labeling. Synthetic spectra broaden the diversity of training data by modeling biochemical variability and measurement effects, providing a scalable framework for pretraining and model evaluation. We explore transformer-based models and alternative spectral encoding schemes to identify representations that transfer effectively across diverse analytical and clinical tasks.

We aim to establish a generalized foundation model for blood ¹H NMR spectra with applications including metabolite quantification, biomarker discovery, and personalized health assessment.

Sophia Hoyer
German Aerospace Center (DLR)
Assessing spatial scale effects in multi-sensor fire fuel mapping in the heterogeneous landscapes of Tasmania

Sophia Hoyer, Anke Fluhrer, Florian Hellwig, Steve Harwin, Jukka Matthias Krisp, Thomas Jagdhuber

Effective fire-risk management in Tasmania requires vegetation maps that capture both broad fuel patterns and small, highly flammable gorse (Ulex europaeus) infestations. Yet gorse often occurs in fragmented patches that disappear in coarser land-cover products, raising uncertainty about how much fuel information is lost when regional maps are produced at moderate or coarse resolution. To clarify these scale effects we compare Object-Based Image Analysis-Random Forest fuel mapping at 0.5, 3 and 10 m in a heterogeneous Tasmanian agricultural landscape using fused optical, LiDAR and SAR features. Beyond accuracy at each scale, we quantify how classes merge, disappear, or persist between resolutions using transfer matrices and analyse how large a gorse patch must be to remain detectable at coarser scales. F1 scores are consistently high across scales (76%–99%), yet class-level behaviour differs substantially. The 3 m model achieves the highest gorse classification performance while maintaining geometric coherence of these shrub patches. When transferred from 0.5 m, 76% of fine-scale gorse area remains represented at 3 m, compared to only 36.8% at 10 m. Detection probability at 3 m increases monotonically with patch size, whereas at 10 m even large patches (10,000–30,000 m) are detected in only 60% of the cases. These results demonstrate that high within-scale accuracy does not guarantee cross-scale persistence of fine-grained fuels. 3 m resolution provides optimal scale–patch alignment for regional fuel-zone delineation in Tasmania, whereas sub-metre imagery is required for explicit identification of individual gorse infestations. Overall, the results confirm that spatial aggregation disproportionately affects narrow and fragmented vegetation types. Resolution choice is therefore not merely a technical setting, but a decisive factor in whether hazardous fine fuels remain visible in regional fuel assessments.

Yifan Xia
University of Applied Sciences Weihenstephan-Triesdorf 
Beyond soil maps - AI to generate soil geoinformation

Yifan Xia, Michael Blaschek, and Mareike Ließ
State Authority for Geology, Mineral Resources and Mining (LGRB), Regional Council Freiburg

Soil function evaluation increasingly demands spatially explicit, depth-resolved soil geoinformation that conventional soil maps deliver only as generalized map units with uniform property values. This project develops a data science approach based on pedogenesis to generate 3D soil geoinformation for the state of Baden-Wuerttemberg. It is a mathematical optimization approach that uses machine learning to condense the soil parameter space into a limited number of spatial soil units. The objective function couples two machine learning tasks in a single optimization. Unsupervised learning groups soil profiles by their similarity, measured on the soil properties relevant to the target soil functions, into soil units with the lowest unit-internal soil variability and the highest inter-unit difference. Supervised learning links these units to the soil-forming environment by predicting them from covariates such as climate, relief and parent material derived from remote sensing, digital terrain analysis and others. Both tasks are optimized jointly, searching for the optimal number of units that are well explained by their environment. Each resulting unit is defined by a multivariate parameter distribution along the depth profile including uncertainty. The approach is first developed on a small high-quality reference dataset and then scaled to a much larger legacy soil profile database covering different land use types. The final 3D soil data product of multiple physical and chemical soil properties can then be used for state-wide soil function evaluation.

Abel Barreto
Institut für Zuckerrübenforschung 
AI-assisted weed control in sugar beet: How practical is field robotics?

Abel Barreto, Dirk Koops, Jonathan Eggers, Stefan Paulus, Anne-Katrin Mahlein
Institute of Sugar Beet Research (IfZ), 77 Holtenser Landstrasse, DE - 37079 Göttingen, Germany

Field robots have the potential to fundamentally transform weed management in sugarbeet production. Their performance is primarily evaluated based on weed control efficacy, crop damage, and herbicide savings. Although modern systems already achieve levels of weed control comparable to conventional approaches, further optimization is still required, particularly regarding reliable plant classification, reduction of crop injury, and determination of optimal application timing. In two field experiments conducted in Sieboldshausen, Germany, during 2024-2025, two commercial robotic systems FarmDroid FD20 and Farming-GT Series 2024 were evaluated using high-resolution UAV imagery and visual field assessments. The results demonstrate that, in addition to weed control efficacy, crop plant losses represent a critical factor influencing overall system performance. GNSS-based systems offer advantages in this regard, as they can identify crop locations without repeated plant classification, whereas camera-based AI systems carry a higher risk of misclassification and erroneous decisions. Early growth stages (BBCH 10–12) proved particularly critical for both robotic platforms, as crop damage occurring at these stages is largely irreversible. At later growth stages (BBCH 14–16), plants may partially recover; however, recovery is often associated with yield penalties. Evaluated robotic systems enabled a reduction in plant protection product use from 75 up to100% while maintaining comparable crop stand density. The findings highlight that each robotic system possesses a specific operational profile and must be integrated individually into cropping systems. Conventional herbicide strategies, such as broadcast post-emergence applications, can only be transferred to robotic systems to a limited extent. Instead, effective deployment requires management practices tailored to the respective technologies, including sensing systems, mechanical weeding components, and spot-application approaches. Overall, autonomous field robots show considerable potential as a sustainable alternative to conventional weed control. Key challenges remain the large-scale implementation under practical farming conditions and the targeted integration into existing production systems to fully realize both ecological and economic benefits.

Josef Eiglsperger
TUM School of Life Science
LiveSen-MAP | NutriSen: Fertilization based on your plants‘ needs. But now way faster with the help of Machine Learning!

Believing in a new, optimized era of fertilization through understanding nutrient concentrations in plant sap, we provide smart fertilization recommendations by combining on-site real-time nutrient sensor measurements, remote sensing, and machine learning (ML). Aiming to save 2 million tons of fertilizers per year, 2 billion € in fertilizer imports, and 20 million tons of CO2 in Europe, we yet had to conquer one of the biggest obstacles which has prevented us from implementing this into modern-day agriculture. A farmer participating in previous field trials summed it up as follows: “It (the sensor) has to analyze much […] faster if this is to be implemented in practice; otherwise, with 40 wheat fields to measure, you won’t have time for anything else”. Our novel real-time nutrient-sensing device had a response time of one hour, which was impractical for industrial applications. But the development of an efficient ML algorithm enabled us to drastically reduce response time. Additionally, our lightweight approach is executable on a microcontroller and, therefore, directly employable on-device. Our user-centric approach allows the farmer to select the preferred response time based on the desired error rate (varying from a five-minute measurement with a 6% expected error rate to a 30-minute measurement with a 1% expected error rate). With this final piece of the puzzle, we can further strive for our overarching goal to reduce fertilizer use (20% on average, saving around 80€/ha€), increase crop yields (up to 10%, generating an extra 240€/ha), and build resilience and independence.

Prof. Dr. Thorsten Schmidt
University of Applied Sciences Weihenstephan-Triesdorf 
Decoding RNA Properties with Machine Learning and Large Language Models: From Sequence Grammar to microRNA Function

RNA molecules carry biological information far beyond their primary sequence. Their structure, dynamics, interactions and regulatory context determine how they function in living systems. Recent advances in machine learning and large language models are now enabling computational methods to learn RNA “grammars” directly from large sequence datasets and to connect these patterns with structural and functional properties.

This talk will briefly introduce recent developments in RNA property prediction, including deep-learning and RNA language-model approaches for predicting structure, molecular interactions and functional classes. A particular focus will be placed on microRNAs, short non-coding RNAs that regulate gene expression through context-dependent interactions with target transcripts. While classical microRNA target prediction relied mainly on seed matching, conservation and thermodynamic features, modern AI-based approaches increasingly capture non-canonical binding, sequence context and transcript-level regulatory effects.

The talk will highlight how these methods can support hypothesis generation in RNA biology, from target prediction to regulatory network analysis and disease association studies. At the same time, it will discuss current limitations, including data bias, incomplete ground truth, interpretability and the need for experimental validation.