Resource-Constrained Federated Learning for Precision Agriculture Using ESP32 Edge Devices

Abstract

This paper presents a resource-efficient federated learning framework for precision agriculture, designed for deployment on resource-constrained solar-powered ESP32 edge devices. The proposed system enables privacy-preserving, decentralized training on local sensor data such as soil moisture, temperature, and humidity without transmitting raw data to the cloud. Model updates are aggregated using the Federated Averaging algorithm, and the resulting global model is quantized to support low-power microcontroller deployment. Experiments conducted on a real-world smart farming dataset demonstrate 99% classification accuracy in crop disease detection after ten federated communication rounds, with a communication overhead of approximately 0.06 MB per round. The results validate the feasibility of integrating federated learning with renewable energy-powered off-grid edge computing, enabling scalable, lowlatency, and sustainable agricultural intelligence.