Federated Learning for data-sovereign training of weed and crop detection machine learning models

As digitalization progresses, AI applications are increasingly finding their way into various agricultural settings. These solutions have great potential, but the underlying AI models require a lot of diverse and realistic data for accurate predictions. This could be realized by centrally collecting data from many farms and attempting to build highly generalized models from them. However, centralized data collection raises privacy concerns for farmers since they want to protect their company-specific information. Federated Learning (FL) offers a solution by training models locally and sharing only model parameters, preserving data sovereignty during training. This paper presents experiments on weed and crop detection systems, demonstrating the potential of FL in agriculture. We first reproduce experiments from literature using YOLOv8 to benchmark against traditional centralized models. We then apply FedAvg to compare FL with centralized training in homogenous and heterogenous cases. Promising results show FL maintains performance while safeguarding data privacy, even across different datasets and locations. The current limitations of our approach will be discussed, as well as aspects of future work.