SIP1
Visual scouting for disease detection, yield estimation,
and precision spraying in vineyards
Sector
Viticulture
Country
Greece
SIP Leader
Eden Library
Participants
7 farmers + 1 advisor
The challenge
Conventional scouting in Greek vineyards is slow, labour-intensive, and inconsistent. Disease is often found days or weeks after it takes hold, and spraying follows the same calendar rhythm for the whole vineyard, regardless of where the problem actually is.
This leads to unnecessary pesticide use, higher costs, and missed yield potential. SIP1 set out to replace guesswork with continuous, real-time data from the vine itself, using computer vision mounted on equipment already in use.
How the Viewer works
The Viewer is a camera and computing unit mounted on a tractor as it moves through vineyard rows. It captures continuous high-resolution video of the vine canopy, processing it in real time or uploading it for cloud analysis, depending on which model is deployed.
As the tractor passes through the rows, AI models detect disease, assess canopy density and vine vigor, and count visible grape clusters. All detections are mapped to GPS coordinates, building a live picture of the vineyard that updates with every pass.
The output is a set of digital maps: where disease is present, canopy density row by row, vine vigor zones, and estimated fruit count. Vineyard managers access these through the dashboard and can generate targeted spray prescriptions, treating only what needs treatment.
Tractor moves through rows. The Viewer captures continuous canopy video.
AI models detect disease, estimate canopy density, vigor zones, and fruit count.
Tractor moves through rows. The Viewer captures continuous canopy video.
AI models detect disease, estimate canopy density, vigor zones, and fruit count.
Two deployment models tested
CLOUD MODEL
Captures full video and GPS data, then uploads via Wi-Fi for remote processing. Delivers the highest accuracy on disease maps and canopy profiling.
Best for: post-session analysis, high connectivity environments
Processes video in real time directly on the tractor. Transmits only lightweight results via 3G. Works in remote vineyards with no Wi-Fi coverage.
Best for: real-time field decisions, remote vineyards, low-connectivity environments
What farmers said
After Year 1, participating farmers and advisors shared their experience, adoption intentions, and concerns about the Viewer system.
Ease of use and adoption
The Cloud mode is perceived as easy to use, requiring only moderate digital skills and integrating well with existing farm machinery. It is strongly endorsed for wider adoption and seen as a potential contributor to regional economic growth.
Data privacy
Farmers raised concerns about data privacy, particularly the transfer of farm data off-site. The Edge mode is preferred in this regard, as it keeps data on-site and helps reduce digital inequality.
Recommended approach
Users favour a phased approach: start with the Cloud mode to support data collection and AI model development, then transition to Edge deployment for improved efficiency, lower energy use, and sustained on-farm operation.
“Start with Cloud. Transition to Edge.”
Year 1 headline results
faster detection
Cloud returns the first detection in 11 to 17 seconds. Both models then run continuously at 10 FPS.
water saved
Targeted spraying versus conventional blanket application across all vineyard rows.
disease
Both cloud and edge models detected every confirmed disease case. Zero missed.
Year 1 results: full breakdown
Results are organised across 4 dimensions. Each compares the Cloud and Edge deployment models side by side. Dimension 4, is represented above in the section What farmers said.
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Time to first detection | 11–17 s (foliage 17s, disease 12s, fruit 11s) | 19–95 s (foliage 95s, disease 23s, fruit 19s) | Cloud up to 82% faster at first result. Both deliver continuous 10 FPS processing thereafter. |
| Disease detection accuracy | 100% recall, 33% precision — 16 false positives | 100% recall, 53% precision — 7 false positives | Both models catch every real disease. Edge produces less noise and fewer unnecessary alerts. |
| Yield estimation | −10% under-count (472 counted vs 527 actual) | +46% over-count (770 counted vs 527 actual) | Yield estimation is functional but needs improvement. Cloud underestimates; Edge overcounts. Calibration is a priority for Year 2. |
| Canopy coverage (LWA%) | 46.9% measured vs 51.9% ground truth | 38.7% measured vs 51.9% ground truth | Cloud matches ground truth more closely, providing a stronger basis for precision spray prescriptions. |
| Water savings | Edge: 40% water reduction via targeted spraying. Cloud: no real-time spraying applied in Year 1. | Targeted spraying reduces water, pesticide use and runoff. Cloud prescription maps planned for Year 2. | |
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Data transmitted per session | 1.26 GB via Wi-Fi (~25 min upload) | 1.74 MB via 3G (under 2 min) | Edge transmits 700× less data and works where Wi-Fi is unavailable. |
| Field power consumption | 474 W total (capture, storage, processing, maps) | 57 W in the field (inference and results only) | Edge uses 88% less power in the field. Heavy processing stays server-side. |
| Equipment fuel use | 2.8 L/km — identical for both models | AI integration adds no mechanical overhead. Tractor fuel economy unchanged. | |
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Data formats | MKV/JPG, CSV, YAML, GeoJSON maps (yield, disease, spray, vigor) | CSV, YAML, GeoJSON maps (yield, disease, spray, vigor) | Both use open standard formats. GeoJSON maps work directly in farm management tools. |
| Connectivity | Wi-Fi (high bandwidth required for raw image upload) | 3G cellular (lightweight results only) | Edge suited to remote areas. Cloud needs reliable field connectivity. |
| Software portability | Hardware-agnostic. Deployable on any cloud or on-premise server. | Platform-dependent. Migration requires significant redevelopment. | Cloud more portable across infrastructure. Edge optimised for specific EDEN hardware. |
| Farm data sharing | Raw imagery and GPS kept private. Not shared with third parties. | Processed outputs in open formats. Shareable with external tools. | Edge data is more openly interoperable. Cloud protects raw farm imagery. |
| Hardware mounting | Identical for both — suction cups, magnetic mounts, or roll-bar plate. Tractor-agnostic. | Flexible across tractor models. No permanent modification required. | |
OPENAGRI OS-BASED SERVICES USED
The following OpenAgri OS-based services are integrated in SIP1:
Weather Data Service
Digital Farm Calendar
Reporting Service
PILOT DETAILS
SIP Leader: Eden Library
Scientific partner: AUA
Third parties: Tselepos winery, Kyr-Yianni winery
Sector: Viticulture / Vineyards
Country: Greece
Participants: 7 farmers + 1 advisor