This project aims to identify unknown correlations among input features that affect and display crop development during the vegetation period and thus crop yield. The focus is set on the interactions between crop and plant stress such as diseases, nutrient deficiencies, toxicities or drought stress. Doing so, we are looking for a link between different growth phases and stress influence on yield development.
For a detailed and precise prediction of plant disease occurrence and spatio-temporal dynamics we integrate information from crop stand monitoring and environmental modelling. We reduce the data input to the most relevant features to develop a highly accurate model for defining new crop protection thresholds and for decision making. On one site, an optimal timing should be approached, on the other side, the potential of site specific or even single crop application of plant protection measures is investigated. Hereby especially the aspect scale is considered. This approach is of benefit for tailored and specific plant protection management strategies, but also of high interest for disease resistance breeding by a better understanding and prediction of plant stress reactions.
Symptoms of nutrient deficiencies visible to the human eye often appear when plants are already severely damaged. Thus, we propose to systematically dissect a variety of early, relevant nutrient deficiencies as well as toxicities and drought stress. We establish nutrient deficiency and drought stress experiments and monitor the crop nutritional status. These data are used to develop deep learning algorithms that are able to predict specific nutrient deficiencies and separate them from other stressors.
Furthermore, we generate virtual realities of plant development in different environments like soil or climate and for different plant traits such as root distribution or leaf shape. For that we are using functional-structural plant models (FSPMs). The FSPMs predict the architecture of the plant, its state and the plant development for different environments and plant properties. This virtual dataset can subsequently be used to train neural networks that link observations to plants states and future plant development. Some advantages of using a simulation model to train machine learning systems compared to training them directly on input-output datasets, are for example that more data can be generated to train the model and the generated data are based on the current process understanding which reduces the number of possible input-output relations to those that are physically and biologically possible. In addition, a reality check of the simulated “virtual realities” is important. This leads to new insights on the interpretation of sensor data, as well as support for decision making in practical agriculture or plant breeding.