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

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

Problem Statement
The paper addresses the critical need for precision agriculture in Nepal, a nation where approximately 66% of the population relies on farming. Current agricultural practices are largely subsistence-based and depend on traditional experience or basic meteorological forecasts, which often fail to account for rapid climate changes and extreme weather events. Existing decision-support systems frequently suffer from a lack of integration between dynamic weather data and static soil properties, or they fail to model the complex spatio-temporal dependencies inherent in Nepal's diverse geographical landscape, which encompasses 1,359 distinct agricultural locations. Furthermore, farmers often lack access to timely, localized information regarding pest control, subsidies, and crop management.

Methodology
The authors propose a unified system comprising three core components: weather prediction, crop recommendation, and a query-answering interface.

1. Data Collection and Preprocessing:

   • Location Data: Manually collected data for 1,359 agricultural regions, including longitude, latitude, altitude, and names.
   • Weather Data: A 42-year historical dataset from NASA covering 1,359 locations. Key variables (Temperature, Precipitation, Humidity, etc.) were preprocessed via correlation testing, standardization, and the creation of temporal snapshots with a 120-day lag.
   • Soil Data: Retrieved via API, including pH, organic matter, nitrogen, and other nutrients.
   • Crop Data: Curated datasets detailing optimal growth conditions (pH, nutrients, weather ranges) for various vegetables.
   • Graph Construction: Locations were modeled as nodes. Edges were defined based on geodesic distance (<15 km) or altitude difference (<50 m), with weights inversely proportional to these differences.

2. Weather Prediction Models:
   Two deep learning architectures were trained to forecast 30-day weather parameters (humidity, precipitation, temperature) using 100 days of historical data:

   • Spatio-Temporal Graph Convolutional Network (STGCN): A Graph Neural Network (GNN) that propagates information between neighboring nodes to capture spatial dependencies, extended to process sequences of graphs for temporal trends.
   • Transformer-Based Graph Model: A hybrid architecture combining temporal convolution (short-term patterns), a temporal transformer block (long-term dependencies), and graph convolution (spatial dependencies).
   • Training: Models were trained using supervised learning with an Adam optimizer, 70/15/15 data splits, and evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

3. Crop Recommendation Engine:
   The system integrates predicted weather data with static soil attributes. A scoring algorithm calculates a suitability score for each crop at a specific location using an inverse distance metric:
   $$score = \frac{1}{1 + |location\_data[feature] - mid\_val|}$$
   where $mid\_val$ is the midpoint of a crop's ideal range for a specific feature. The engine outputs a ranked list of recommended crops.

4. Retrieval-Augmented Generation (RAG) Chatbot:
   A conversational agent was developed to answer farmer queries regarding crops, pests, and subsidies. It employs a hybrid retrieval approach using SPLADE (for sparse representations) and dense-sparse vectors to query a knowledge base stored in Pinecone. The system generates context-sensitive answers, primarily in the Nepali language, using Cohere's API.

5. Deployment:
   The entire system is deployed via a mobile application built with React Native, featuring a Django backend and SQLite3 database.

Key Results

• Weather Prediction: The STGCN model outperformed the Transformer-based model, achieving a Mean Squared Error (MSE) of 0.011 compared to 0.013 for the Transformer. The authors attribute the STGCN's superior performance to its explicit modeling of interactions between geographical locations, which is crucial for regional weather patterns.
• System Usability: The mobile application successfully provides location-based crop recommendations and real-time updates. User feedback indicated the system is usable and relevant, particularly in rural settings where personalized guidance is scarce.
• RAG Performance: Qualitative assessments showed the RAG system consistently generated relevant and helpful responses to diverse agricultural queries, effectively grounding answers in the provided knowledge base.

Significance and Claims
The paper claims that this research bridges the gap between advanced machine learning and practical agricultural application in Nepal by:

• Integrating Data Sources: Successfully unifying spatio-temporal weather predictions with static soil data within a single framework, addressing a limitation in prior studies that often treated these factors in isolation.
• Modeling Complexity: Demonstrating that Graph Neural Networks, specifically STGCNs, are more effective than standalone Transformer models for capturing the spatial and temporal dependencies required for accurate regional weather forecasting in geographically diverse areas.
• Empowering Farmers: Providing a mobile-first solution that delivers actionable, localized insights and conversational support, thereby promoting informed decision-making, potentially increasing crop yields, and enhancing resilience to climate variability.

Limitations and Future Work
The authors acknowledge that the system relies on historical data that may be inconsistent and lacks comprehensive quantitative metrics specifically for crop recommendation accuracy. They note that the models may not fully capture highly complex environmental interactions or specific microclimates. Future work is proposed to include real-time satellite data, develop robust quantitative evaluation metrics for recommendations, enhance model architectures, and improve mobile accessibility with offline capabilities.