Crop Recommendation and Agricultural Query Answering System Using Spatio-Temporal Graph Neural Networks and Hybrid Retrieval Augmentation

Researchers developed an integrated agricultural system combining Spatio-Temporal Graph Convolutional Networks for weather forecasting, machine learning-based crop recommendations, and a retrieval-augmented generation chatbot to support precision farming in Nepal. The STGCN model achieved superior accuracy in 30-day weather predictions across 1,359 locations, enabling localized crop suggestions matched to soil properties and climate conditions.

This research demonstrates the practical application of advanced machine learning architectures to solve real-world agricultural challenges in resource-constrained regions. The dual-model approach—using STGCN for weather forecasting and hybrid retrieval systems for farmer support—represents a thoughtful engineering solution that prioritizes accuracy alongside accessibility. The STGCN's performance advantage over transformer-based models (MSE 0.011 vs. 0.013) validates the importance of explicitly modeling spatial dependencies in climate data, where neighboring geographical regions significantly influence local weather patterns.

The broader context reflects growing recognition that AI systems must integrate domain expertise with local data to deliver meaningful value in agriculture. Climate variability poses increasing threats to crop yields globally, and precision agriculture tools have historically remained inaccessible to smallholder farmers due to cost and technical barriers. This open-source approach to combining weather prediction, agronomic matching algorithms, and conversational AI addresses that gap directly.

The market implications extend beyond individual farmers to agricultural extension services, agritech companies, and climate adaptation initiatives in developing nations. Mobile deployment signals commercialization potential, particularly in South Asian markets where agricultural productivity gains directly impact food security and rural incomes. The verified user feedback from rural settings provides credibility that the system solves genuine problems rather than optimizing for academic metrics alone.

Future development should focus on scaling model performance across diverse agro-climatic zones, integrating real-time market price data for crop selection, and expanding the agricultural knowledge base to languages beyond English. The system's success depends on sustained partnerships with local agricultural extension networks and farmer cooperatives to ensure continued relevance and data accuracy.

• →Spatio-Temporal GCNs outperformed transformer models in 30-day weather forecasting with MSE of 0.011 across 1,359 Nepalese locations.
• →The system combines weather prediction, soil-based crop recommendations, and a RAG chatbot into a single mobile-accessible platform for farmers.
• →Integration of local climate and soil data with machine learning enables personalized agricultural guidance previously unavailable to smallholder farmers.
• →User validation from rural deployments confirms practical usability and relevance, reducing implementation risk for agritech adoption.
• →The approach demonstrates viable pathway for AI systems to improve climate resilience and crop yields in resource-limited agricultural communities.