GNN-ML-FRL: a graph-enhanced meta-adaptive federated learning framework for scalable pest identification and ernvironmental modeling

Abstract

Abstract Heterogeneous agro-ecological factors, insect breeding, and climate change are serious challenges to sustainable agricultural management. The study proposes a graph-enhanced meta-adaptive federated learning framework (GNN-ML-FRL) to address the challenges in precision agriculture. The proposed framework integrates Federated Learning (FL) for collaborative training of models in a decentralized manner across geographically distributed farms, Meta-Learning (ML) for rapid adaptation to changing environmental factors, and Graph Neural Networks (GNNs) for capturing spatial dependencies among agricultural entities. A comprehensive multivariate IoT environmental dataset with 52.56 million time-series observations gathered from 500 dispersed sensors over a 12-month period, the IP102 insect pest recognition benchmark (75,222 images across 102 species), and curated genomic datasets from MaizeGDB and the Rice Annotation Project Database for genotype-informed modeling are the three standardized datasets used to assess the framework. Experimental results show statistically significant improvements ( p  < 0.01) over CNN and graph-based baselines, achieving 89.3% Top-1 accuracy, 7.8% higher generalization performance, and 12.4% reduction in prediction loss across geographically unseen farms. SHAP-based explainability further indicate that environmental accuracy-related features contributed nearly 63% positive influence, while loss-related factors contributed 37% negative influence, validating model robustness. Geographic generality is confirmed by site-out validation using IoT data, and resilience is improved under varied crop conditions by genotype-informed graph modeling. The findings show that a scalable and statistically sound framework for data-driven pest identification and environmental modeling in precision agriculture may be achieved by combining spatial graph reasoning, meta-adaptive learning, and decentralized training.