Field-deployable coffee yield estimation from mobile phone images using branch segmentation and occlusion correction

International audience Accurate coffee yield estimation is critical for crop management, labor and financial planning, and value-chain transparency, including compliance with the EU Deforestation Regulation. However, manual cherry counting remains labor-intensive, error-prone, and unreliable as worker fatigue sets in, highlighting the need for automated, scalable alternatives. This study introduces a novel deep-learning framework for automated coffee cherry counting using images captured with low- to mid-range smartphones across diverse smallholder farming contexts. The pipeline combines automated branch segmentation, cherry detection, and a regression-based correction module to account for occluded cherries, accommodating different data-capture modalities. We evaluated the framework on 7025 annotated images from Colombia, Peru, Honduras, and Uganda, covering both Coffea arabica and Coffea canephora (Robusta) coffee species. Under optimal image-capture conditions (i.e., full background isolation), the model achieved high accuracy, reaching an R2 of up to 0.96 and reducing the Mean Absolute Percentage Error (MAPE) to as low as 10% at the plot level, outperforming state-of-the-art methods. By reducing manual effort and addressing real-world constraints in smallholder settings, this approach offers a strong foundation for scalable coffee yield estimation. Future research should prioritize human-centered design validation and detailed cost-benefit analyses to support widespread adoption and long-term sustainability.