@PhdThesis{diss_mods_00023094,
  author = 	{Hamer, Wolfgang Berengar},
  title = 	{Spatial prediction of the infestation risks of winter wheat by the pathogens Blumeria graminis f. sp. tritici (Powdery mildew) and Puccinia triticina (Brown rust) in Schleswig-Holstein using machine learning techniques},
  year = 	{2018},
  publisher = 	{Christian-Albrechts-Universit{\"a}t zu Kiel},
  address = 	{Kiel},
  keywords = 	{phytopathogenic fungi; plant protection; spatio-temporal prediction; Decision Tree; Random Forest; k-Nearest-Neighbor; phytopathogene Pilze; Pflanzenschutz; r{\"a}umliche und zeitliche Vorhersage; Entscheidungsbaum},
  abstract = 	{Wheat is one of the most important cereals in the world. Phytopathogens such as powdery mildew or brown rust can considerably reduce wheat yields. By treatment with fungicides, infections with these pathogens can be contained. It is of decisive importance to be informed about upcoming dangerous infestation events in real time to be able to respond to them. To define which infestation events are to be classified as yield-relevant, this work uses the damage threshold concept. This concept assumes that the exceedance of a 70 {\%} threshold value for powdery mildew and of a 30 {\%} threshold value for brown rust of infected plants in a field would threaten the yield of the complete stock of winter wheat and, thus, suggests the application of fungicides.

The main objective of this thesis is the spatial prediction of the probability of exceedance above this damage threshold. In order to achieve this goal, a concept was developed that regionalises the hourly weather data on a daily basis and subsequently uses these data as input parameters for predicting the pathogen-specific behaviour. Besides, the modelling concept uses supervised machine learning techniques to generate models based on these aggregated weather data, regionalised climate data and manually collected infestation data. The following learning methods are used to predict the occurrence of infestation spatially: k-Nearest Neighbor, Decision Trees, Boosted Decision Trees and Random Forests. The concept was examined iteratively using various evaluation methods, and thus the pathogen-specific performance of the models was tested concerning the prediction of the probability of infestation. The model results generated with the machine learning methods were then integrated into a web-based prediction system, which provides interested users with the probability of dangerous infestations.},
  url = 	{https://macau.uni-kiel.de/receive/diss_mods_00023094},
  file = 	{:https://macau.uni-kiel.de/servlets/MCRFileNodeServlet/dissertation_derivate_00007706/Hamer_Dissertation_Pathogen_prognosis.pdf:PDF},
  language = 	{en}
}