From Hyperspectral Unmixing to EUDR Compliance: Scalable Cocoa Traceability in West African Agroforestry Systems

Monitoring cocoa farming within complex tropical agroforestry systems remains a significant challenge for Earth Observation, particularly given the EU Deforestation Regulation's requirement for supply chain verification of deforestation-free status at the farm level. In West Africa, distinguishing cocoa trees from forest shade canopies with standard multispectral satellite data is difficult because their spectral signatures are similar.

This study introduces a scalable approach that uses high-resolution hyperspectral imagery from the Wyvern Dragonette satellite constellation to identify cocoa within mixed agroforestry landscapes. The methodology uses Google Earth Engine as the cloud platform and integrates available hyperspectral images, which are limited by frequent cloud cover, with ground-truth data from Abeya’s smallholder supply chain network.

The proposed methodology uses unique spectral 'forest fingerprints' from adjacent native forests to characterise the background canopy. Pixels within farm boundaries that deviate from these forest signatures but correspond to the spectral patterns of known cocoa plantations are identified. These cocoa-specific signatures are subsequently associated with multispectral Sentinel-2 data and pre-trained geospatial foundation models, facilitating cocoa tracking in regions lacking hyperspectral imagery.

This is achieved by utilising the high spectral dimensionality of the Wyvern Dragonette constellation, which captures 31 bands, to resolve sub-pixel mixing between cocoa and forest shade trees that multispectral sensors typically cannot disentangle. These high-fidelity insights are subsequently used to fine-tune pre-trained geospatial foundation models, effectively transferring hyperspectral intelligence to the broader spatial and temporal coverage of the Sentinel-2 archive. This approach demonstrates the potential for emerging satellite constellations to transition from experimental platforms to operational, interdisciplinary monitoring tools that support environmental policy and sustainable supply chain decision-making.

To support validation in data-sparse, smallholder contexts, the framework incorporates participatory field observations when available. Planned farmer questionnaires, aimed at estimating cocoa tree counts and farm-level planting characteristics, will be explored as a complementary source of reference information. These self-reported inputs are intended to provide an independent check on spatial cocoa predictions and help contextualise spectral patterns observed from space. While the availability and completeness of such data may vary, this approach highlights the potential of farmer-generated information to strengthen EO-based monitoring of agroforestry systems.

By focusing on operationally effective modelling choices rather than theoretical optimality, this work outlines a practical pathway for integrating emerging hyperspectral satellite constellations into scalable geospatial workflows. The proposed framework aims to support future assessments of cocoa traceability, EUDR compliance, and sustainable land-use monitoring in tropical agroforestry systems.