Abstract:
Problem: Smallholder farmers in Sri Lanka struggle with foliar diseases in climate-sensitive
crops like tomatoes and bananas under erratic weather, leading to late interventions, heavy
losses, and pesticide overuse. No existing system combines image-based detection with
forward-looking weather forecasts and actionable advice in Sinhala or Tamil. Agricultural
experts can review and approve recommendations when farmers have doubts.
Methodology: SmartAgri is a web-based platform that applies convolutional neural networks
to analyze user-uploaded images of tomato and banana leaves, detecting diseases in under 1.2
seconds. It enriches these detections with real-time 3-day weather forecasts via public APIs to
estimate outbreak risks and generate proactive, location-specific recommendations. A
responsive React + TypeScript frontend with Flask microservices delivers results instantly
through a lightweight, multilingual UI (Sinhala, Tamil, English). By fusing image-based
diagnosis with climate-driven insights, SmartAgri empowers smallholder farmers to reduce
pesticide use and build resilience against erratic weather.
Results: On 1,200 banana and tomato leaf images, the hybrid model achieved 99.78% accuracy,
outperforming EfficientNet-B0 (33.38%) and matching MobileNetV2. Inference time was
under 1.2 seconds per image. A 4-week rural pilot maintained over 99.4% uptime. In usability
testing, 12 farmers gave a SUS score of 76 (benchmark: 70), and 11 completed the workflow
without help. Experts rated recommendation relevance 4.6/5 and architecture 4.5/5, confirming
technical strength and practical value.
