Abstract
Côte d'Ivoire and Ghana, the world's largest producers of cocoa, account for two-thirds of global production. Despite its economic importance, precise maps of cocoa-planted areas have been historically unavailable, hindering sustainability governance and accurate quantification of forest loss. This study utilizes a deep learning framework combined with multi-spectral satellite imagery and canopy height data to create high-resolution (10m) maps of cocoa plantations for both countries. Validated in situ, the results indicate that cocoa cultivation is a primary driver of deforestation, associated with over 37% of forest loss in protected areas in Côte d’Ivoire and 13% in Ghana. Furthermore, official reports are found to substantially underestimate planted areas by up to 40% in Ghana.
The Governance Gap in Cocoa Production
The cocoa industry in West Africa is characterized by millions of smallholder farmers operating on small plots, often under forest canopies. This fragmentation makes traditional mapping difficult, leading to a significant data deficit. Without accurate maps, verifying "deforestation-free" claims, monitoring illegal encroachment into forest reserves, and accurately estimating yields for economic planning remains nearly impossible. This "governance gap" limits the effectiveness of sustainability initiatives and agricultural extension services.
Methodology: Deep Learning and Multi-Modal Data
To address this, researchers developed a deep learning model to translate satellite data into country-scale probability maps.
Data Integration
The model leverages nine spectral bands from Copernicus Sentinel-2 imagery (2018–2021). A critical innovation is the inclusion of a canopy height map derived from GEDI (Global Ecosystem Dynamics Investigation) spaceborne LiDAR data. This structural information helps the model distinguish cocoa trees (typically ~8m tall) from taller primary forests or shorter ground crops.
Model Architecture and Validation
An ensemble of ten neural networks was used to produce continuous probability maps, allowing for uncertainty quantification. The training set included over 100,000 georeferenced cocoa farm polygons. To ensure real-world accuracy, the team conducted a three-month ground campaign in Côte d’Ivoire, visiting over 2,000 locations for in situ validation. The resulting model achieved an overall accuracy of 88.7%.
Key Findings: Deforestation and Underreporting
The study reveals a stark contrast between official statistics and ground reality.
Underestimation of Planted Area
Official FAOSTAT figures underestimate Ghana’s cocoa area by nearly 40% (1.63 Mha reported vs. 2.71 Mha detected). This suggests significant production occurs in unrecorded or recently cleared regions, such as the Volta region.
Encroachment into Protected Areas
The maps provide clear evidence of massive encroachment into "classified forests." In some areas, such as the Niegre forest in Côte d’Ivoire, cocoa now covers over 80% of the land. Overall, cocoa is associated with 37.4% of forest loss in Côte d’Ivoire’s protected areas and 13.5% in Ghana’s since 2000.
The Yield Paradox
By providing more accurate area measurements, the study recalculated cocoa yields. In Ghana, the mean annual yield was found to be ~320 kg/ha, significantly lower than previous estimates of 400–530 kg/ha. This suggests that earlier figures were biased toward high-performing farms, masking widespread low productivity and soil depletion.
Conclusion and Future Outlook
These high-resolution maps serve as a vital tool for conservation and economic development. While the current method relies on temporal aggregation to overcome cloud cover, future iterations could incorporate Synthetic Aperture Radar (SAR) for more frequent updates. Furthermore, adapting this model to other regions like Southeast Asia will require additional training data to account for different agroforestry practices. Despite these challenges, the study provides a transparent, data-driven foundation for improving the sustainability of the global cocoa supply chain.
Summary updated