Decentralized Federated Learning with Adaptive Model Compression for Precision Agriculture IoT Networks in Sub-Saharan Africa

Precision agriculture holds tremendous potential for enhancing crop yields and food security across Sub-Saharan Africa, yet deploying machine learning systems in this region confronts fundamental barriers including limited network connectivity, constrained computational resources, and pressing concerns over agricultural data sovereignty. This paper presents AgriFL-Edge, a decentralized federated learning framework purpose-built for resource-constrained agricultural Internet of Things networks prevalent in Sub-Saharan African farming communities. The proposed framework advances three principal contributions: (i) a Bandwidth-Aware Dynamic Sparsi cation algorithm that adjusts compression ratios based on real-time bandwidth availability and model convergence trajectory, (ii) a fault-tolerant gossip protocol that eliminates reliance on centralized aggregation servers while gracefully handling prolonged network partitions, and (iii) AgriNet-Lite, a lightweight convolutional neural network architecture optimized for crop disease detection on microcontroller-class edge devices. Rigorous experiments conducted through real-world deployments in Nigeria, Ghana, and Kenya demonstrate that AgriFL-Edge attains 94.2% disease classi cation accuracy while reducing communication overhead by 73% relativeto standard federated averaging. The framework operates successfully on devices with less than 512 MB RAM and tolerates connectivity interruptions exceeding 48 hours without appreciable model degradation. Field trials engaging 847 smallholder farmers across 12 agricultural cooperatives validate practical feasibility and sustained adoption under authentic operational conditions.