Optimizing Soil-Based Crop Recommendations with Federated Learning on Raspberry Pi Edge Computing Nodes

Publication Type

Conference Proceeding

Publication Date (Issue Year)

2024

Journal Name

International Conference on Internet of Things: Systems, Management and Security

Abstract

The need for sustainable agriculture and food security necessitates adopting data-driven methods in crop management. Traditionally, centralized approaches make informed decisions based on extensive data from field-deployed sensor networks. However, transmitting vast amounts of data from connected devices to the cloud is often impractical due to connectivity constraints and limited computational resources. These challenges are especially critical when optimizing agricultural productivity through soil-based crop recommendation models. Federated Learning (FL) presents a viable solution by decentralizing model training across edge devices like Raspberry Pi clients. This work explores the integration of precision agriculture and edge computing, demonstrating how FL supports soil-driven crop recommendations. We conducted experiments using a publicly accessible crop recommendation dataset to evaluate the impact of FL parameters, including the number of clients, training rounds, and local epochs, on predictive accuracy, computational efficiency, and communication overhead. Our results demonstrate that the FL setup achieved a high testing accuracy of 92%, significantly outperforming previous studies. Memory usage during training ranged from 1.5 to 1.9 GB, confirming the feasibility of FL on resource-constrained devices without risking memory overflow. Communication overhead analysis revealed efficient data handling with balanced load distribution. These findings highlight the scalability of FL and the importance of managing client participation and training rounds to handle communication overhead, particularly in bandwidth-limited environments or areas with intermittent connectivity, such as remote agricultural fields. The results underscore the potential of FL at the edge to enhance resource utilization and enable localized decision-making in precision agriculture, making it a practical solution for sustainable crop management.

Keywords

Optimizing Soil-Based, Crop Recommendations, Federated Learning, Raspberry Pi Edge Computing Nodes

Rsif Scholar Name

Rehema Hamis Mwawado

Rsif Scholar Nationality

Tanzania

Cohort

Cohort 4

Thematic Area

ICTs Including Big Data and Artificial Intelligence

Africa Host University (AHU)

University of Rwanda (UR), Rwanda

Share

COinS