Federated learning with dual dynamic quantization optimization in smart agriculture

Abstract

The convergence of Federated Learning (FL), Machine Learning (ML), and the Internet of Things (IoT) creates promising opportunities for smart agriculture, where connectivity constraints and limited device resources pose major bottlenecks. To address these challenges, we propose a Dual Dynamic Quantization Optimization (FedDDO) framework that jointly integrates quantizer design, adaptive bit allocation, and quantization-error-aware aggregation. On the client side, FedDDO dynamically adjusts quantization bit-widths according to real-time resource conditions, while on the server side, aggregation weights are optimized based on quantization error feedback. A novel Minimum Relative Quantization Error (MRQE) quantizer is designed to align with unbiased error assumptions, and theoretical analysis under non-convex settings provides convergence guarantees. Extensive experiments on both standard benchmarks (CIFAR-10/100) and agriculture-specific datasets (rice seedling classification and disease recognition) demonstrate that FedDDO effectively reduces communication costs and accelerates convergence, achieving competitive accuracy while preserving domain applicability.