Applying Federated Learning on Decentralized Smart Farming: A Case Study

In the field of Smart Agriculture, accurate time series forecasting is essential for farmers to gather and evaluate relevant information about various aspects of their work, such as the management of harvests, livestock, crops, water and soil. One commonly used method for trend forecasting in time series is the Long Short Term Memory (LSTM) Recurrent Neural Network (RNN) model, due to its ability to retain context for longer periods and enhance performance in context-intensive tasks. To further improve the results, the use of Federated Learning (FL) can be implemented, allowing multiple data providers to simultaneously train on a shared model while preserving data privacy. In this study, a Centralised Federated Learning System (CFLS) is leveraged, that implements and evaluates the efficacy of FL in smart agriculture through the use of datasets produced by such infrastructures. The system receives data from multiple clients and creates an optimised global model through model federation. Consequently, the federated approach is compared with the conventional local training to explore the potential of FL in real-time forecasting for the Smart Farming sector.