It appears you don't have support to open PDFs in this web browser. To view this file, Open with your PDF reader
Abstract
Agriculture is currently facing a series of pressing challenges such as climate change, soil degradation, and population growth, which pose a significant threat to food security and environmental sustainability. In the past, the green revolution, which relied heavily on the use of inputs and new genotypes, led to significant gains in crop productivity. Yet, the use of intensive inputs is no longer sustainable due to its negative environmental impacts and to the societal concerns. Therefore, it is crucial to develop new solutions that enable continued crop improvement while reducing inputs such as fertilizers, pesticides, and water.
To effectively address these challenges, plant breeding plays a vital role, requiring a comprehensive understanding of plant behavior from biological, physiological, and agronomic perspectives. Large-scale phenotyping is essential to this process, as it involves characterizing plants in diverse environmental and cultural situations, enabling the identification of desirable traits for selection and optimization of crop management practices. While manual phenotyping is labor-intensive and subject to human subjectivity, automated phenotyping techniques utilizing sensors are being developed to allow for faster, more precise, and objective plant characterization. However, their application under outdoor conditions remains limited due to constraints such as wind, changing light conditions and dense canopy.
In this work, a mobile phenotyping platform has been developed, equipped with a set of cameras. The platform has been used to study several winter wheat trials using two RGB cameras and one multispectral camera. The current thesis focuses mainly on the development of the image analysis pipeline, with a pronounced investigation of the use of artificial intelligence algorithms. The methods have enabled the detection of wheat ears to count the density per hectare, the detection of disease-related damage, and the estimation of biophysical variables such as above-ground biomass, leaf area index and nitrogen content. The deep learning approaches showed to be better as traditional machine learning methods and tends to better generalized. Estimated throughout the growing seasons, these traits were used as predictors of grain yield and yield components providing a deeper understanding of these complex yet highly relevant traits.
Despite significant advancements in the field of automated phenotyping, its adoption by end-users such as breeders is still very limited. Therefore, it is essential to continue raising awareness among stakeholders in the agricultural industry about the benefits of using these technologies to improve sustainable agricultural production. By embracing new technologies, we can help to develop crops that are better adapted to diverse conditions, increase crop productivity, and reduce the negative impacts of agriculture on the environment.