Künstliche Intelligenz und Data Science

Projektteam:
 Viacheslav Barkov (Ph.D. Student), Prof. Dr. Martin Atzmüller (UOS), Dr. agr. Robin Gebbers (ATB)

As the volume and complexity of natural system sensor data grow, an increasing need emerges for advanced predictive modeling, knowledge discovery, and explainable decision-making in environmental monitoring contexts. This research develops advanced machine learning approaches to estimate environmental parameters from increasingly complex sensor data. By focusing on precision, interpretability, and robustness, the project spans multiple sensing modalities from proximal soil sensors to remote laser scanning of forests. In pedometrics and soil science, the work advances tabular regression and deep learning models for digital soil mapping. For forestry applications, the research develops deep learning methods for processing laser scanning point cloud data to estimate forest plantation biometrics. Across both domains, the research emphasizes predictive precision, robust uncertainty quantification, and interpretability of results. This integrated approach contributes to improved environmental parameter estimation, supporting more reliable agricultural and environmental decision-making.

Projektteam:
 Moritz Lucas (Ph.D. Student), Prof. Dr. Björn Waske (UOS), Dr.-Ing. Ralf Pecenka (ATB)

Trees outside of forests (TOF) play a critical role in biodiversity, cultural heritage, and landscape connectivity, especially in intensive agricultural areas and agroforestry systems. However, these trees are often neglected in conventional monitoring approaches, which tend to focus on forests or urban trees. This project aims to develop a robust methodology to map and classify TOFs, providing researchers, ecologists, and policymakers with better insight into their spatial distribution and structural characteristics. By integrating multisensor remote sensing data with state-of-the-art deep learning techniques, this project seeks to create a scalable tool for improving landscape-level assessments and supporting evidence-based decision-making.

Projektteam:
 Clara Lößl (Ph.D. Student), Dr.-Ing. Ralf Pecenka (ATB), Dr. Thomas Jarmer (UOS), Prof. Dr. Kai-Uwe Kühnberger (UOS), Prof. Dr. Martin Atzmüller (UOS), Prof. Dr. Tim Römer (UOS)

Accurate and efficient tree monitoring is increasingly needed for comprehensive forest inventory, structural assessment, and carbon stock estimation. Moreover, the growing demand for sustainable energy production drives the need for efficient monitoring of fast-growing tree plantations, such as Short Rotation Coppices used for biomass and bioenergy. Machine learning and geometric deep learning methods for analyzing LiDAR-derived 3D point clouds offer promising approaches for these applications. This project encompasses the complete pipeline from field data acquisition using aerial laser scanning (ALS) to method development, requiring interdisciplinary expertise bridging agricultural sciences and computer science. The aim is to investigate, explain, and improve machine learning approaches, with a focus on modern geometric deep learning architectures that learn directly from 3D point cloud representations. The research addresses individual tree segmentation from large-scale, high-resolution point clouds and automated regression of biometric parameters such as trunk diameter at breast height (DBH) and above-ground biomass (AGB). Both self-collected plantation data and publicly available forestry benchmark datasets are utilized to ensure generalizability. The outcomes are expected to enhance inventory practices, reduce labor-intensive manual measurements, and contribute to sustainable forest management, biodiversity conservation, and climate change mitigation efforts.

Projektteam:
 Freddy Sikouonme (Ph.D. Student), Prof. Dr. Martin Atzmüller (UOS), Prof. Dr. Marina Höhne (ATB)

AI models can be trained on a variety of data, such as sensor data, satellite data, genomic data, image data, weather data, and climate data. This can be used in the context of different applications and several use cases in the bioeconomy domain, e.g., for improving performance regarding disease detection on plants, precise plant monitoring, smart water use, and optimizing waste recycling processes. Such models can then help, for example, to detect pests, identify weeds, and optimize the use of fertilizers and pesticides. Although those AI models learn complex relationships between different components in the bioeconomy, those relations typically remain hidden within the often called black box model, leading to no or only poor knowledge about the underlying connections of the driving factors in the input data. Here, the field of explainable AI (XAI) has emerged, with the aim to provide insights into the decision-making strategies and to showcase the important features and complex relationships the decision is based on. In this project, methods of explainable AI for multimodal data will be developed for explaining AI models trained on multimodal data with applications in the bioeconomy. With the novel XAI methods we aim to showcase the complex relationships between different types of data, such as genetics, soil composition, weather, and climate data, learned by the AI model. We aim to contribute to the generation of novel knowledge that tackles the enormous challenge of an efficient realization of the bioeconomy.