Automated Detection of Vine Plants and Missing Vines in the Saint-Émilion Vineyard
Automated Detection and Geolocation of Vine Stocks and Missing Plants Across Seven Vineyard Plots: Results Derived from Photogrammetric Processing and Deep Learning Approaches Applied to 8,500 Aerial Images Acquired by R3Drone. Post-Processing and Image Analysis Conducted by Drones Imaging.
The wine industry, historically rooted in traditional practices, is currently undergoing a major transformation driven by advances in image acquisition and analysis technologies. Automated vine stock detection using Deep Learning, combined with high-resolution UAV-based aerial imagery, enables fine-scale, spatially explicit vineyard management and represents a key step toward the automation and robotization of viticultural operations.
The operational workflow begins with multi-overlap aerial image acquisition using an unmanned aerial vehicle (UAV). On average, approximately one hectare can be covered within five minutes of flight time, generating several hundred images. With a sub-centimeter ground sampling distance (GSD < 1 cm), these images allow for highly accurate representation of the individual morphological characteristics of each vine stock.
The acquired data are subsequently processed using photogrammetric techniques to generate a high-resolution georeferenced orthomosaic. Based on this dataset, a Deep Learning model specifically trained for vine detection is applied. The algorithm leverages visual descriptors such as morphology, texture, radiometric signature, shadow patterns, and spatial context to automatically identify each plant.
The analysis resulted in the mapping of 44,423 vine stocks and the detection of 568 missing plants, geolocated with an estimated planimetric accuracy on the order of 10 cm.
Automated Mapping and Detection of 44,423 Vine Stocks and 568 Missing Plants
The integration of UAV-based photogrammetry and Deep Learning methods applied to very high-resolution orthophotos provides several key structural advantages for precision viticulture:
Spatial Accuracy and Operational Performance
Deep learning models enable robust vine stock detection with decimetric planimetric accuracy. Automated processing allows rapid coverage of large areas, significantly optimizing acquisition and analysis timelines compared to conventional methods.Optimization of Human Resources and Cost Reduction
Automated detection replaces manual counting operations, which are time-consuming and subject to variability. This approach reduces labor requirements while improving reproducibility and objectivity of results.Generation of High-Precision Cartographic Reference Data
The production of a comprehensive georeferenced map of vine stocks and missing plants establishes a spatial reference framework that can be leveraged for:
agricultural machinery guidance using GNSS RTK,
intra-parcel robotic operations,
targeted replanting (gap filling) interventions.
GIS Integration and Advanced Spatial Analysis
The generated data (georeferenced vector layers) can be directly integrated into a Geographic Information System (GIS). This enables detailed spatial analyses, including planting density assessment, distribution of missing plants, intra-parcel variability, and correlation with soil, topographic, or agronomic datasets.Temporal Monitoring and Individual Traceability
Each detected vine can be assigned a unique identifier within a geospatial database. This structure supports multi-year individual monitoring (seasonal yields, plant health status, water stress, diseases, fungal infections), facilitating differentiated and historically informed management at the vine scale.
TECHNICAL SPECIFICATIONS
- Geomatics & image analysis : Drones Imaging.
- UAV flights and aerial data acquisition : R3Drone.
- Spatial accuracy : decimetric.
- Number of images processed : 8500.
- Total surveyed area : 7ha.
- Number of mapped vine stocks & missing : 44423 + 568 manquants.
- UAV platform : DJI M300.
- Photogrammetry software : Agisoft Metashape.