CropGuard: An AI-Driven Agricultural Diagnosis and Intelligent Recommendation System for Smallholder Farmers

CropGuard is an AI agricultural diagnosis tool designed specifically for smallholder farmers. It combines the Vision Transformer visual model with the lightweight large language model Qwen2.5 to enable crop disease identification and recommendation of localized organic treatment solutions. The system adopts an edge computing architecture to support offline operation, aiming to provide low-cost, accessible services for smallholder farmers in resource-constrained areas and support sustainable agricultural development.

Global pests and diseases cause 20-40% of crop yield losses annually. Smallholder farmers struggle to accurately identify and handle diseases due to lack of professional knowledge and resources, while traditional agricultural consulting services have limited coverage and high costs. CropGuard emerged to address deployment needs in resource-constrained environments. Through edge computing, it ensures normal operation even in areas without stable networks, providing low-cost and accessible diagnostic services.

Dual-Model Collaboration Architecture: The visual diagnosis module uses Vision Transformer (ViT) to identify crop leaf diseases, and transfer learning improves accuracy with limited labeled data. The intelligent recommendation module adopts the lightweight large language model Qwen2.5 7B GGUF to generate localized treatment solutions that meet organic agriculture standards.

Edge Computing Optimization: Model quantization and compression reduce memory and computing requirements, supporting operation on ordinary smartphones/edge devices. Core functions are fully offline, adapting to areas with weak networks. Visual model pruning and distillation reduce computational load, enabling real-time image processing.

Intelligent Disease Identification: Taking a photo of the leaf allows disease identification, including symptom feature analysis, crop type recognition, severity assessment, and differentiation of similar diseases.

Localized Treatment Solutions: Organic and environmentally friendly solutions are prioritized, considering local climate/crops/resources, and low-cost prevention measures and preventive suggestions are provided.

User-Friendly Design: Supports multilingual and voice interaction, with a built-in offline knowledge base to eliminate language and cultural barriers.

Visual Model Training: The ViT model is trained on datasets covering multiple crops, different growth stages/severity levels/shooting conditions, and data augmentation is used to improve robustness.

Language Model Fine-Tuning: Qwen2.5 7B is fine-tuned using data such as agricultural manuals, organic cases, expert experience, and farmer Q&A to output professional and easy-to-understand solutions.

Mobile Deployment: Android/iOS apps are provided, with a hybrid architecture of native and model inference to ensure a smooth experience.

• Food Security: Reduce yield losses, increase self-sufficiency, and ensure regional food security.
• Sustainable Agriculture: Promote organic solutions, reduce chemical pesticide use, and protect the ecological environment.
• Empowering Smallholder Farmers: Lower technical barriers, enable farmers in remote areas to access professional guidance, and narrow the digital divide.
• Knowledge Inheritance: Combine traditional agricultural experience with AI technology to drive agricultural innovation.

In the future, we will expand crop coverage, integrate meteorological data to provide disease early warning, establish farmer communities for knowledge sharing, connect with government agricultural subsidy and technical support projects, and continue to promote the inclusive application of AI technology to serve social needs.