Hierarchical federated learning cuts energy costs for plant disease AI

A new paper accepted at the 2026 ERAS Conference tackles a key challenge in precision agriculture: how to train deep learning models for plant disease detection across distributed IoT sensors without draining energy budgets. The team from multiple European research institutes proposes hierarchical federated learning (HFL), which inserts intermediate aggregation layers between edge clients and a central server. This reduces both communication overhead and computational load compared to standard federated learning.

Testing three convolutional neural networks (EfficientNet-B0, ResNet-50, MobileNetV3-Large) with three aggregation strategies (FedAvg, FedProx, FedAvgM), the authors systematically measure trade-offs between classification accuracy and energy consumption. They also introduce a power- and energy-aware optimization framework that lets deployers select the best model-aggregator combination for specific constraints. Results show that careful configuration can maintain high diagnostic accuracy while slashing energy use, making large-scale agricultural AI more viable.