Smart agriculture, driven by the Internet of Things (IoT) and artificial intelligence, generates vast amounts of data from distributed sensors and devices on farms. Traditional centralised machine learning approaches struggle in rural settings due to intermittent connectivity, limited bandwidth, and data privacy concerns. Federated learning (FL) has emerged as a promising approach to address these challenges by collaboratively training models directly on edge devices, for example farm sensors and edge computers, without sending raw data off-site. This thesis presents the design, development, and evaluation of a standards-driven, hardware-based federated learning testbed for smart agriculture. The testbed consists of six NVIDIA Jetson Nano edge computing nodes. A primary objective was to evaluate and compare modern, open-source FL frameworks in a realistic edge environment. Therefore, we specifically tested the FLIGHT framework. FLIGHT was selected for its notable features, including a Function-as-a-Service (FaaS) architecture facilitating serverless deployment and native support for hierarchical topologies relevant to distributed farm networks. We developed a reproducible methodology for deploying and benchmarking FLIGHT on these resource-constrained devices, using the Fashion-MNIST image classification dataset as a proxy for agricultural sensor data to compare performance against known benchmarks and simulation expectations. Key contributions include: 1) the physical testbed itself; 2) an open and repeatable experimental framework centered around FLIGHT; and 3) a comprehensive performance analysis under realistic network conditions and device constraints, providing data to bridge the gap between simulation results and practical deployment challenges. The results demonstrate that the federated model achieves competitive accuracy while significantly reducing raw data transfer. Findings indicate that careful configuration of FL can mitigate the impact of limited connectivity, and that even low-power devices can collaboratively train useful models within reasonable time-frames. These insights validate the viability of FL in smart farming scenarios. The developed testbed and its accompanying benchmarking methodology lay a foundation for future research and deployment of FL in agriculture, bridging the gap between theoretical simulations and real-world farm deployments. This work’s significance lies in providing both a practical tool for researchers to rigorously evaluate FL strategies in edge environments and guidance for stakeholders aiming to deploy privacy-preserving AI in agriculture at scale.