AI-powered cocoa mapping: Sub-metre resolution barely beats decametric with embeddings

A new arXiv paper from researchers at the European Commission’s Joint Research Centre and partners evaluates whether sub-metre resolution satellite imagery is truly necessary for accurate cocoa mapping in heterogeneous smallholder landscapes. The study compared very high resolution (VHR) 0.5m Pleiades imagery against decametric inputs: a 10m Sentinel-2 annual composite, TESSERA embeddings, and AlphaEarth Foundations (AEF). They also tested four existing operational products. Using a landscape-stratified accuracy assessment with 2,821 independent reference points across gradients of tree cover and fragmentation, the VHR model achieved the highest performance (F1=0.92) and maintained F1>0.90 across all strata. Among decametric inputs, TESSERA performed best (F1=0.86), followed by AEF (F1=0.82) and Sentinel-2 alone (F1=0.76). The Kalischek product led existing maps at F1=0.83, comparable to the internally trained AEF model.

Performance differences between VHR and decametric approaches increased with landscape fragmentation and under low/high tree cover density. This suggests targeted VHR acquisition is most beneficial in complex cocoa landscapes, while foundation-model embeddings provide a scalable alternative for large-area mapping. The findings have direct implications for deforestation monitoring, supply-chain transparency, and regulatory compliance under schemes like the EU Deforestation Regulation. The study underscores that foundation models like TESSERA can nearly match sub-metre accuracy without the cost of VHR imagery, making wide-area cocoa surveillance more practical and accessible.