Deep learning-based estimation of Kent mango yield by commercial export category under real field conditions in...

Manual visual counting remains the standard procedure for estimating Kent mango (Mangifera indica L.) yield in northern Peru, an inconsistent practice that compromises harvest planning, transport dimensioning and commercial negotiation. Existing computer vision approaches to mango detection resolve fruit-versus-background discrimination but rarely classify fruit into the commercial categories that determine its market value. This study developed and validated a deep learning method that simultaneously detects and classifies Kent mango fruit into three export categories (air freight, sea freight and reject) from RGB images acquired under real field conditions in the Motupe district, Lambayeque. A purpose-built dataset of 400 photographs was compiled from 200 trees across ten commercial farms using two opposite views per tree; 6,544 patches of 1,280 × 1,280 pixels with 20% overlap were generated, containing 66,465 manually annotated instances. Five architectures were trained under identical experimental conditions: YOLOv8m, YOLO11m, YOLO26m, RT-DETR-l and Mango-YOLO, the latter a proposed architecture built on YOLO26m through the incorporation of a CBAM attention module, a BiFPN neck and the MPDIoU regression loss. Mango-YOLO achieved the best performance across all five metrics, with a precision of 0.930, a recall of 0.940, an F1-score of 0.930, a mAP@0.5 of 0.940 and a mAP@0.5:0.95 of 0.730, at an inference time of 114.94 ms per image. Per-category analysis revealed heterogeneous behaviour, with F1-scores of 0.960, 0.845 and 0.792 for sea freight, air freight and reject respectively. The selected model was integrated into a three-tier web application that scales fruit counts from sampled trees to the whole farm and converts them into kilograms through Monte Carlo simulation with category mean weights, returning a 90% interval. Validation against harvested production recorded on the processing-plant scale across the ten farms yielded a mean absolute percentage error of 2.63% and a coefficient of determination of 0.9993, outperforming the field supervisor's manual estimate by a factor of 3.3.