Towards Livestock Location Prediction in Rural Areas using Federated Learning and Edge Computation

Due to increasing presence of technology in agriculture, livestock management has become increasingly important and new methods are developing rapidly. However, rural areas make it extremely hard to track livestock and due to challenges such as limited connectivity and lack of centralized storage.The inability to transfer large volumes of data in a consistent manner makes the usage of a centralized data storage highly unfeasible, making it hard to effectively train models and distribute them. In addition to this, the lack of powerful centralized processing solutions further complicates the training and inference of models.Existing solutions of livestock tracking methodologies mostly use non-machine learning models or centralized deep learning solutions. The issue with those methods is that they both have different drawbacks. Non-machine learning methods, such as ARIMA-family models, have difficulty capturing some more complex and dynamic pattern movements of the livestock. As for the centralized deep learning method, in spite of being much more powerful, have a need for large datasets and their transfer to a central server. This creates a significant challenge in rural areas due to many constraints such as bandwidth limitations.In order to get by these limitations, this thesis proposes an approach utilizing Federated Learning (FL) in combination with Edge Computing for rural livestock position prediction.FL gives the ability to multiple local devices (clients) to train a shared global model in collaboration without a need to send data to a centralized server. Instead of collecting data, only model weights are shared, which as a result significantly reduces communication costs.Empirically the results show that the performance of the FL-based approach is superior.The approach achieved an average position prediction accuracy of approximately 52meters after only five rounds of server-client communication. In comparison, this shows a significant improvement with accuracy increasing almost 50% from the best non-ML approach, ARIMAX (101m), and and additional increase of over 10% in comparison to a centralized deep learning approach. The results show that FL is able to give high accuracy predictions while eliminating the need for centralized storage and addressing the problem of limited connectivity.