CocoaDeep: A preliminary study of the performance sensitivity to datasets of Faster RCNN, YOLO and transformer networks for cocoa pod detection.

Authors

Philippe Borianne, Frédéric Théveny, Llorenç Cabrera-Bosquet, Sabine-Karen Lammoglia

Published in

PloS one. Volume 21. Issue 7. Pages e0351657. Epub Jul 22, 2026.

Abstract

Farmers must be able to estimate their crop yields at various growth stages for effective management of their farms and to enable them to interact with cooperatives or traders as early as possible. Here we developed an AI-based cocoa pod detection method using low resolution colour images of cocoa trees on farms in Côte d'Ivoire. We compared nano and extra-large architectures of six neural networks, including Faster RCNN, Baidu's Real-Time Detection Transformer (RTDetr), Detr-ResNet Vision Transformer (ViT), YOLOv5, YOLOv8 and YOLOv11. These networks were trained with 7,850 annotated cocoa pods on 400 low resolution images, and validated in two independent datasets: a 42 low resolution images containing 990 annotated pods, and a 100 low resolution images containing 2,400 annotated pods. The performances of the nano YOLOv8 and YOLOv11 networks were 2% higher than that of the RTDetr networks and 5% higher than that of the YOLOv5, ViT and Faster RCNN networks with an F1-score of 77% on all images and up to 90% on foreground trees. The dominance of nano architectures suggests that the extra-large architectures, which contain 20-30-times more neurons, may not have been fully trained. The study of learning performance curves showed that extra-large networks were unable to outperform nano networks, which contradicts the theory. After review, the annotated dataset was found to contain inconsistencies. The inconsistency of the training and validation data and their limited quantity restricted the objectivity of comparisons between network architectures. Finally, although the average detection performance of RTDetr for cocoa pods was only 2% lower than that of the YOLOv8 network, it was definitively excluded from the candidate models because its per-image processing time was 15-20% higher than that of YOLOv8 and YOLOv11. However, with a performance sensitivity to data of less than 0.5%, YOLOv8 Nano became the best option.

PMID:
42485275
Bibliographic data and abstract were imported from PubMed on 23 Jul 2026.

Read full publication at:
Please sign in to see all details.

Stats

1-terrible, 9-excellent. How would you rate this publication? Sign in in to submit your rating.

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.