Günder, Maurice: Knowledge-Driven Intelligent Systems for Precision Agriculture : A Demonstration on Disease Spread Prediction. - Bonn, 2026. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-90795
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-90795
@phdthesis{handle:20.500.11811/14348,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-90795,
doi: https://doi.org/10.48565/bonndoc-929,
author = {{Maurice Günder}},
title = {Knowledge-Driven Intelligent Systems for Precision Agriculture : A Demonstration on Disease Spread Prediction},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2026,
month = aug,
note = {Organic cultivation, resource efficiency, and sustainability are gaining increasing importance in times of climate change. To address these issues, modern agriculture is increasingly relying on the use of drones to monitor and analyze crop stands. They enable non-invasive and rapid collection of large-scale imagery and geodata. A logical step in this development is the processing of these data through the integration of artificial intelligence (AI). Purely data-driven approaches, however, quickly reach their limits when only small amounts of annotated data are available or the collection of additional (meta) data is associated with high personnel and financial effort.
In agriculture, decades of research have already produced extensive expert knowledge in a wide range of disciplines. This knowledge can be used to improve the performance of data-driven AI models and increase their applicability in practice. This dissertation aims to show how data-driven models can be combined with knowledge in order to address important agricultural challenges, while at the same time remaining generally applicable.
As a use case, the spread of the leaf spot disease Cercospora in sugar beet is examined, from data collection with drones, through the prediction of infestation severity, to epidemiological modeling. The first building block is the acquisition of an extensive image database from the georeferenced drone recordings. Using robust algorithms, a semi-automated workflow is developed to detect plant positions in temporally separated drone images of sugar beet fields. The result is a spatially and temporally referenced database of individual plant images, which serves as the basis for the second building block. In this second step, the image data is used, together with environmental data and assessments by human experts, to train an AI model to predict the severity of Cercospora infestation. The third building block uses this model to estimate the time of infection of individual plants. In combination with meteorological data and epidemiological knowledge, an infection model is developed that can predict the spatio-temporal spread of Cercospora in sugar beet stands. This could enable the agriculture of the future to take targeted and needs-based measures to combat plant diseases and thus minimize the use of plant protection products.
The AI models developed are in many respects based on the transformer architecture, which in recent years has achieved partly disruptive advances in various domains. When used deliberately, they enable high interpretability and realistic modeling of the underlying processes. A particular focus of this dissertation is on model interpretability. Through application-oriented modeling, models can be designed in such a way that their predictions are comprehensible to domain experts. This is an important step toward increasing trust in AI-based systems and promoting their acceptance in practice.},
url = {https://hdl.handle.net/20.500.11811/14348}
}
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-90795,
doi: https://doi.org/10.48565/bonndoc-929,
author = {{Maurice Günder}},
title = {Knowledge-Driven Intelligent Systems for Precision Agriculture : A Demonstration on Disease Spread Prediction},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2026,
month = aug,
note = {Organic cultivation, resource efficiency, and sustainability are gaining increasing importance in times of climate change. To address these issues, modern agriculture is increasingly relying on the use of drones to monitor and analyze crop stands. They enable non-invasive and rapid collection of large-scale imagery and geodata. A logical step in this development is the processing of these data through the integration of artificial intelligence (AI). Purely data-driven approaches, however, quickly reach their limits when only small amounts of annotated data are available or the collection of additional (meta) data is associated with high personnel and financial effort.
In agriculture, decades of research have already produced extensive expert knowledge in a wide range of disciplines. This knowledge can be used to improve the performance of data-driven AI models and increase their applicability in practice. This dissertation aims to show how data-driven models can be combined with knowledge in order to address important agricultural challenges, while at the same time remaining generally applicable.
As a use case, the spread of the leaf spot disease Cercospora in sugar beet is examined, from data collection with drones, through the prediction of infestation severity, to epidemiological modeling. The first building block is the acquisition of an extensive image database from the georeferenced drone recordings. Using robust algorithms, a semi-automated workflow is developed to detect plant positions in temporally separated drone images of sugar beet fields. The result is a spatially and temporally referenced database of individual plant images, which serves as the basis for the second building block. In this second step, the image data is used, together with environmental data and assessments by human experts, to train an AI model to predict the severity of Cercospora infestation. The third building block uses this model to estimate the time of infection of individual plants. In combination with meteorological data and epidemiological knowledge, an infection model is developed that can predict the spatio-temporal spread of Cercospora in sugar beet stands. This could enable the agriculture of the future to take targeted and needs-based measures to combat plant diseases and thus minimize the use of plant protection products.
The AI models developed are in many respects based on the transformer architecture, which in recent years has achieved partly disruptive advances in various domains. When used deliberately, they enable high interpretability and realistic modeling of the underlying processes. A particular focus of this dissertation is on model interpretability. Through application-oriented modeling, models can be designed in such a way that their predictions are comprehensible to domain experts. This is an important step toward increasing trust in AI-based systems and promoting their acceptance in practice.},
url = {https://hdl.handle.net/20.500.11811/14348}
}





