SIP3
Drone-based AI detection of Colorado Potato
Beetle for targeted pest control
Sector
Arable / Potatoes
Country
Belgium
SIP Leader
EV ILVO
Farmer partner
ISP
The challenge
Potato farmers in Belgium apply pesticides to their fields twice every season as a standard practice, regardless of whether the Colorado Potato Beetle is actually present. This blanket approach increases costs, accelerates pesticide resistance, and risks contaminating ground and surface water.
SIP3 set out to replace this routine with precision: autonomous drone scouting that detects exactly where the beetle is, at what growth stage, and in what density — so treatment is targeted only where it is needed.
How the system works
An autonomous UAV flies a pre-planned route over the potato field, capturing high-resolution images of the crop canopy. As it flies, an AI detection mode analyses the imagery to identify Colorado Potato Beetle and larvae at different growth stages, locating infestations with spatial precision.
The system generates infestation maps showing exactly which parts of the field are affected and how severely. These maps are then converted into prescription maps that guide targeted pesticide application — treating only the zones where the pest is confirmed present.
The system was tested in two configurations: a Cloud variant where images are uploaded to a remote server for processing; and an Edge variant where detection happens directly on the drone during the flight, with only compact results transmitted to the cloud.
UAV autonomously surveys the potato field, capturing high-resolution crop imagery.
AI mode detects Colorado Potato Beetle at different life stages, mapping infestation zones.
Infestation map generated. Prescription map created for targeted spraying.
Farmer applies pesticide only in affected zones. Unaffected areas left untreated.
Two deployment modes tested
CLOUD MODE
Images are uploaded from the drone to a remote server for processing. Delivers slightly higher detection accuracy and enables centralized data management. Requires stable uplink connectivity during or after the flight.
Best for: areas with reliable connectivity, farms where centralized data management is a priority
EDGE MODE
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: remote fields, low-connectivity environments, real-time field decisions
What farmers said
After Year 1, participating farmers and advisors shared their experience, adoption intentions, and concerns about the Viewer system.
Easy to adopt, no disruption to existing routines
The Edge variant is rated as very easy to use and faces no adoption resistance. The Cloud variant also met initial expectations and showed low resistance. Both modes are seen as complementary tools that support existing agronomic practice -- not as replacements that disrupt it.
Connectivity shapes the choice
The Cloud variant is constrained by its dependence on stable high-bandwidth connectivity - a significant limitation in rural Belgian fields. The Edge variant eliminates this dependency by performing detection on the drone itself, making it reliable in low-connectivity environments.
Edge strongly preferred -- and reduces the digital divide
Users strongly favour the Edge variant on both technical and practical grounds. Local processing keeps data on-site, reduces bandwidth and energy use, and ensures reliable operation. The Edge variant is also perceived as improving accessibility and reducing the digital divide -- the Cloud variant was found to be poorly compatible with existing machinery infrastructure.
“Detect only where needed. Spray only where needed.”
Year 1 headline results
Edge completes a full detection cycle in 1.2 seconds. Cloud takes 27 seconds. Both are accurate — Edge is simply faster.
Edge reduces the cost per drone mission from approximately 127 EUR to 94 EUR by eliminating high-volume image uploads.
Cloud: mAP50 0.701. Edge: mAP50 0.687. The difference is small — the beetle is visually distinct enough that even lightweight modes detect it reliably.
Drone propulsion dominates total energy use. Optimising flight paths matters more than computing efficiency for reducing the mission energy footprint.
Year 1 results: full breakdown
Detailed performance metrics comparing Cloud and Edge deployment across three dimensions. Each section can be expanded for full data. Performance graphs for computing and network metrics are available in the technical deliverable — confirm embedding approach with the developer.
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Time to detection (TTD) | ~40s over stable Wi-Fi. Over 4G/3G, upload time for full-resolution images can add minutes. If no uplink during flight, results only available after landing. | ~1.2s. Only compact result payloads transmitted — TTD is predictable and low regardless of connectivity. | Edge delivers near-instant detection. Cloud latency is heavily network-dependent — in rural fields with poor 4G, this is a significant practical limitation. |
| ML mode accuracy | Precision: 0.76 / Recall: 0.674 / mAP50: 0.701 / mAP50-95: 0.257 | Precision: 0.71 / Recall: 0.675 / mAP50: 0.687 / mAP50-95: 0.251 | Cloud offers marginally higher accuracy. The gap is small because the Colorado Potato Beetle is visually distinct — even lightweight modes capture it reliably. For real-time field deployment, Edge efficiency outweighs the marginal accuracy gain. |
| Production cost per mission | Approximately 127 EUR. Communications (15 EUR) and cloud compute/storage (20 EUR) are the main variable costs on top of fixed pilot, transport and drone costs. | Approximately 94 EUR. Communications (2 EUR) and cloud compute/storage (2 EUR) are minimal — only compact results transmitted. | Edge saves approximately 26% per mission. The cost of the scouting mission should be weighed against avoided full-field spray costs — a single avoided application can offset several monitoring flights. |
Note: production cost figures are illustrative benchmarks. Actual costs vary by region, telecom tariffs, and cloud pricing modes. Recalculate with local data for operational planning.
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Computing resources | Cloud server (MSI NUC) sustains 35–37W during inference plateau. CPU-intensive — 60–90% load during the inference phase. Scaling to simultaneous detections requires significant infrastructure provisioning. | Raspberry Pi 5 fluctuates 6–13W. AI accelerator handles inference with minimal CPU load. Full detection cycle: 1.2s. Suitable for real-time deployment on resource-constrained hardware. | Edge completes detection 22x faster (1.2s vs 27s) by offloading inference to a dedicated AI accelerator. Cloud places sustained load on server infrastructure. |
| Network utilisation | Sustained ~25,000 kB/s transmit rate during image upload over Wi-Fi. Cloud mode is heavily dependent on reliable, high-bandwidth uplink. Poor connectivity causes mission delays or data loss. | Brief low-volume spikes only — compact result payloads (bounding boxes, confidence scores, timestamps). Highly resilient to poor connectivity. Cost-efficient for cellular deployments. | Edge sends only results, not images. Near-zero bandwidth dependency — a major advantage in rural and remote fields. |
| Energy for data processing and transmission | Cloud server: sustained 35–37W plateau during inference. Raspberry Pi (drone-side): 5–11W during image capture and upload. | Raspberry Pi: 6–13W. AI accelerator handles inference efficiently. Cloud server receives only lightweight results — draws 10–14W for a short duration with no sustained plateau. | Edge keeps the server in a lightweight supporting role. Note: cloud server in these tests did not use GPU acceleration — incorporating GPU would reduce inference time and energy consumption. |
| Total energy per mission | Same for Cloud and Edge. DJI M350 RTK with dual TB65 batteries: 522 Wh capacity. Typical 10-min flight at 550W = 91.7 Wh. Drone propulsion dominates: 500–600W during flight. A 10-minute mission consumes 90–100 Wh — far exceeding computing and transmission energy combined. Optimising flight paths has greater impact on sustainability than improving computing efficiency. | ||
| Metric | Cloud | Edge | What this means |
|---|---|---|---|
| Data formats | Identical for Cloud and Edge. JPEG input. InfluxDB and Postgres for intermediate storage. Prescription maps in ISOXML, shapefiles, or GeoTIFF — directly usable by precision spraying equipment without conversion. | Consistency of formats is the key interoperability feature. | |
| Networking standards | JPEG uploads to S3 via HTTPS/TLS. Postgres via TCP. REST API for image and result registration. Encrypted and authenticated. High bandwidth required. | Local S3 storage. Only compact results sent to Postgres via REST API. Token-based authentication, encrypted channels. Resilient to low bandwidth. | Both encrypted and authenticated. Edge leverages communication protocols far more efficiently — transmits metadata only, not images. |
| Software switchability | Same modular design for Cloud and Edge. Modular micro-service architecture in both modes. YOLO model, Streamlit, Postgres and InfluxDB are all swappable via REST APIs. High switchability across both variants. | Any component can be replaced without rewriting the pipeline. | |
| Farm data switchability | Same for Cloud and Edge. JPEG input accepted from any drone, camera, or sensor. Prescription map outputs in ISOXML, shapefiles, GeoTIFF — compatible with any compliant spraying equipment or farm management platform. | Input flexibility and output universality are built into the standard. | |
| Hardware switchability | Same for Cloud and Edge. Input: any drone or camera producing JPEG. Edge inference: Raspberry Pi, Jetson, or equivalent AI accelerator. Cloud inference: any CPU or GPU server. Output: prescription maps compatible with any compliant spraying equipment. | Full hardware independence — farms are not locked to a single vendor. | |
OPENAGRI OS-BASED SERVICES USED
The following OpenAgri OS-based services are integrated in SIP3:
Weather Data Service (real-time and historical climate data for infestation risk prediction)
Pest Management Platform (centralized database for logging, analysing, and mapping pest outbreaks)
Digital Farm Calendar (logs UAV scouting sessions and tracks infestation patterns over time)
PILOT DETAILS
SIP Leader: EV ILVO
Farmer partner: ISP
Sector: Arable / Potatoes
Country: Belgium
Target group: Farmers and advisors (co-creation)