Accurate weed detection plays a crucial role in sustainable crop management, enabling targeted interventions that reduce chemical use and improve yield. Deep learning-based weed detection and targeted, site-specific weed management offer significant opportunities for resource and environmental conservation in agriculture. However, current state-of-the-art (SOTA) deep learning models are typically trained on narrow, task-specific datasets and perform poorly when exposed to new, unseen field conditions. This is due to the inherently heterogeneous nature of weed detection data, which for instance varies across crop types, growth stages, soil conditions, imaging modalities, and environmental factors. These models also often suffer from being highly data-hungry, requiring large volumes of labeled data to function effectively. Collecting such data in agriculture is costly and time-consuming. Therefore, there is a pressing need for more data-efficient and generalizable approaches.
In this regard, we aim to explore and validate the hypothesis that pretrained foundation models, which have learned general visual features from large and diverse datasets, can be fine-tuned for weed detection with significantly less labeled data, while maintaining or improving performance compared to traditional SOTA models. These models are hypothesized to be less sensitive to data heterogeneity and capable of generalizing well, thereby enabling broader applicability in real-world agricultural conditions.
Towards data efficient and robust weed detection leveraging foundational models.
MITGLIED IM KOLLEG
seit
Verbundkolleg Life Sciences und Grüne Technologien
Prof. Dr. agr. Heinz Bernhardt
Forschungsgebiet:
- Technik in den Agrarwissenschaften
- systemorientierter Ansatz der Interaktion von Technik, Technologie und Verfahren
Forschungsschwerpunkte:
- sensorgeregelte Präzisionslandwirtschaft
- Agrarlogistik
- Robotik
- Elektromobilität
- Melktechnik
- agrarisches Energiemanagement
Betreute Projekte:
- Vorhersage der Akzeptanz bei Einführung eines On-Farm-Energiemanagementsystems im automatisierten Milchviehstall unter Berücksichtigung demographischer und sozioökonomischer Auswirkungen auf Gesellschaft und den ländlichen Raum
- Grundlagen zur Implementierung eines On-Farm Energiemanagementsystems im Milchviehstall
- Towards data efficient and robust weed detection leveraging foundational models.
Prof. Dr. Florian Haselbeck
Die Professur Smart Farming an der Hochschule Weihenstephan-Triesdorf unter der Leitung von Florian Haselbeck forscht zu KI-basierten Methoden für diverse Anwendungsbereiche im Kontext Nachhaltigkeit. Der Schwerpunkt liegt auf der Entwicklung von Computer Vision- und Time Series Forecasting-Methoden sowie multimodalen Machine Learning-Verfahren, um in enger Zusammenarbeit mit Domänenexpert:innen praxisrelevante und wissenschaftlich fundierte Lösungen zu entwickeln.
Betreutes Projekt:
Towards data efficient and robust weed detection leveraging foundational models.
Harshavardhan Subramanian
Hochschule für angewandte Wissenschaften Weihenstephan-Triesdorf
Master: Statistics and Machine Learning, Linköping University, Sweden
Bachelor: Industrial Engineering and Management, Visvesvaraya Technological University, India
Publication: Bukas, Christina, Harshavardhan Subramanian, Fenja See, Carina Steinchen, Ivan Ezhov, Gowtham Boosarpu, Sara
Asgharpour et al. ”MultiOrg: A Multi-rater Organoid-detection Dataset.” Advances in Neural Information Processing
Systems 37 (2024): 95808-95839.
