Future Harvest: The Smart Vineyard

Grapevine diseases such as Black Rot and Esca (Black Measles) cause substantial yield losses and threaten global viticulture. This paper presents Future Harvest, a real-time mobile disease detection system that integrates YOLOv8s-based object detection, Grad-CAM-based visual explainability, and augmented reality (AR) overlays for interactive field diagnostics. The YOLOv8s model was fine-tuned using transfer learning on a dataset of approximately 9,000 annotated grape leaf images, split into 70% training, 20% validation, and 10% testing. Biologically constrained augmentations preserved disease morphology, and a novel HSV-based surgical masking technique was applied to Grad-CAM outputs to prevent heatmap bleed into healthy tissue. The system architecture includes a Flutter mobile client and a Python Flask inference server, enabling seamless deployment across devices. On the test set, the model achieved [email protected] of 99.34%, [email protected]:0.95 of approximately 91.2%, precision of 96.2%, and recall of 94.8%, with GPU inference latency between 50–100 ms on RTX 30/40 series hardware and increased trust in predictions due to Grad-CAM overlays. These results demonstrate that Future Harvest delivers accurate, interpretable, and scalable disease detection, with practical impact validated through on-farm deployment and user engagement.