A Hybrid RAG Architecture for Context-Aware Sugarcane Irrigation Scheduling

This research presents a Hybrid Retrieval-Augmented Generation (H-RAG) architecture for analyzing IoT time-series data to support irrigation decision-making in sugarcane, a major Thai cash crop that is sensitive to environmental variability. The study integrates five consecutive years (2019–2024) of IoT sensor data—rainfall, temperature, relative humidity, solar radiation, and multi-depth soil moisture—with agronomic knowledge such as Kc, ETc, and the principles of Regulated Deficit Irrigation (RDI), to generate context-aware and explainable recommendations (Explainable AI). The H-RAG model employs hybrid retrieval that combines embedding vectors from time-series data with knowledge graphs to improve retrieval accuracy and reduce misleading outputs. Evaluation is performed against a baseline retrieval, measured by Precision@5, Recall@5, and nDCG, as well as Yield Index, Water Use Efficiency (WUE), and Commercial Cane Sugar (CCS). Experimental results show that H-RAG increases Precision@5 from 0.239 to 0.470 and Recall@5 from 0.317 to 0.470, with XAI coverage rising by more than 25% compared with conventional methods. This reflects the system’s ability to produce context-congruent and traceable recommendations. The study contributes methodologically, technically, and practically to Responsible AI for sustainable water management in agriculture.