AgroSAGE: A Modular Ontology-Driven Framework for Real-Time Sensor Annotation in Smart Agriculture

Global food security issues are growing due to increasing population, farmland decline, and climate instability, raising the need to transition to smart agriculture by utilizing the Internet of Things (IoT) and Artificial Intelligence (AI) for efficient production. Existing IoT frameworks lack domain-specific ontologies, and the processing of high-dimensional sensor data increases cost and latency. AgroSAGE (Smart Annotation and Graph-based Embedding for Agriculture) is proposed. First, agricultural ontologies are constructed using Natural Language Processing (NLP)-based entity extraction and Graph Neural Networks (GNN), capturing crop types, phenology stages, and agronomic actions. Second, a semantic- and ontology-based IoT framework enables protocol reconciliation and semantic mapping across heterogeneous devices. Finally, Dynamic Sparse Principal Component Analysis (D-SPCA) was applied to reduce the number of sensors used while preserving variance and interpretability. The ontology is modularized into crops, sensors, phenology, and action components for lightweight deployment. Real-time annotation was performed at the edges using compressed GNN embeddings and semantic rules. The validation protocol simulated 100–500 devices using MQTT, CoAP, and HTTP. AgroSAGE achieves global sparsity with fewer sensors while maintaining prediction accuracy effectively, resulting in 98% interoperability success, 66% latency reduction, and balanced resource utilization.