Objectives of the service
Cocoa growers and buyers depend on manual field scouting and pod counting to judge crop health and yield. These methods are slow and inconsistent, and cover only a small share of farms. A 2024 peer-reviewed study in Costa Rica found that an established forecasting method overestimated yield by 27–45%. Disease is often found late, and deforestation rules need evidence that is hard to produce by hand.
HiveSight replaces this with continuous, objective monitoring. Scouts are field devices with multispectral and polarimetric cameras and on-device AI that count pods and detect disease on individual trees. Earth observation data extend these sparse scans to whole fields and supply bases, and a causal machine-learning model turns them into yield forecasts with stated uncertainty.
The objective of the project is to define the service with cocoa buyers, evaluate a proof of concept for combining EO and Scout data, and build the business case for deployment. The target is to reduce yield forecast error from ±30–40% towards ±10%, detect disease earlier, and produce audit-ready evidence for the EU Deforestation Regulation (EUDR).
Users and their needs
The users currently targeted are cocoa buyers and the farmers who supply them. The customers involved in the study include the world's largest food and cocoa companies, individual large farmers and farmer cooperatives. The service has already been trialled on six large plantations in Ecuador, and the targeted supply bases are in Ecuador, Brazil, Ghana, Côte d’Ivoire and Indonesia. Their needs are:
-
Yield forecasts across the whole supply base that are more reliable than manual pod counts, with stated uncertainty.
-
Earlier detection of disease, mapped to the affected trees, so that treatment is targeted.
-
Timestamped, geolocated evidence of coverage and land-cover change for EUDR compliance.
-
Lower scouting effort and less variation between scouts.
The main challenges for the project are showing that EO models, calibrated by a small share of scouted fields, remain accurate on fields that are not scouted, and keeping the service affordable per hectare.
Service/ system concept
HiveSight delivers field-level and supply-base-level information: disease maps, pod and flower counts, yield forecasts with uncertainty, and compliance evidence. Customers receive it through the Hive web platform and itemised recommendations.
Scouts are carried through a field and scan a sample of trees. Each Scout carries a multispectral and polarimetric camera, a computer that detects pods and scores them for ripeness, counts flowers and identifies disease lesions on the device, a satellite positioning receiver (GNSS) with real-time kinematic (RTK) correction for precise location, and a GSM modem.
In the cloud, ML models join the scattered tree scans into maps on a 3 × 3 m grid. These maps are updated over time with satellite data in a “living map”. A causal yield model, developed with Causa Ltd, and a forecasting model turn the maps and climate data into yield forecasts. Uncertainty in the maps is used to prioritise which fields the Scouts scan next. Federated learning is designed so that Scout AI models improve without sending raw images.
Space Added Value
Space assets are used in three ways. EO satellites (Sentinel-2 optical, Sentinel-1 and ALOS-2 PALSAR-2 radar), with CHIRPS rainfall, ERA5 climate, canopy height and ESA WorldCover land-cover data, extend Scout observations to every field. Satellite navigation places every scan on the correct tree, and aligns Scout data with satellite pixels and plot boundaries for compliance evidence.
The added value compared with current methods is coverage, connectivity, and objectivity. Manual scouting samples a small share of farms, and satellite-only services lack tree-level disease and pod information. HiveSight combines both: Scouts calibrate the satellite models, and satellites carry the result to unscouted fields. EO data also guide the Scouts, because fields with the highest forecast uncertainty are scanned first.
Satellite communications, from a low-Earth-orbit service, are identified as a potential benefit for more reliable connectivity, but are not part of the current design.
Current Status
The pipeline from Scout to EO fusion to living maps is built. Pod counting and disease identification etc., reach high accuracy. In early tests, EO-based models reach a correlation of about 0.6 for yield and a block-level R² of about 0.7 for disease in cross-validation within scouted fields. Thermal EO data were assessed and found not informative. Scouts carry RTK-capable GNSS and GSM modems.
Large-scale field deployment will begin in January 2027. Starting in Ecuador, then spanning the rest of the cocoa belt.
Prime Contractor(s)
Subcontractor(s)
