GrapeSense-VISION: Vineyard Intelligence System for Interactive Observation and Navigation using Drone-Based Deep Learning

Grape crops are vital for global agriculture, including food security and the wine-making industry. Traditionally, effective vineyard management requires labor-intensive scouting to evaluate ripeness progression and detect plant disease onset, a process that is time-consuming and often catches issues too late for intervention. The GrapeSense project offers a scalable and modern approach to viticulture, leveraging drone-captured imagery and deep learning to automate vineyard monitoring. Working with wineries in California, aerial images were collected across multiple drone flights spanning a complete growing season. These geotagged images are processed through a two-stage deep learning pipeline. First, a YOLOv8 object detection model localizes individual grape clusters and leaves, achieving an mAP50 of 85.0%. Second, specialized transfer-learning classifiers evaluate the localized regions: a ResNet-152 model categorizes grapes into one of seven ripeness stages with 93.28% accuracy, while a fine-tuned ResNet-18 architecture identifies the presence of Black Rot, Esca, or Leaf Blight with 99.61% accuracy. To translate these outputs into actionable insights, this work introduces GrapeSense VISION, an interactive dashboard that aggregates classification results onto a satellite map of the vineyard. Each detection is plotted at its GPS-extracted coordinates with color-coded markers indicating spatial distributions of ripeness levels and disease hotspots. The dashboard also includes temporal filtering by flight date, ripeness distribution charts, and interactive image monitoring views. By intersecting drone-captured data, automated visual analysis, and an accessible interface, GrapeSense provides a practical tool for growers to optimize harvest scheduling and prioritize targeted disease treatments, reducing both labor costs and potential crop-yield loss.