•  
  •  
 

Corresponding Author ORCID Identifer

Mohd Haris Lye https://orcid.org/0000-0001-9248-6607
Marawan Ashraf Eldeib https://orcid.org/0009-0008-5285-424X

Section

Regular Issue

Abstract

Mango cultivation is a vital agricultural sector in many regions, providing significant economic benefits and contributing to environmental sustainability. Traditional mango detection and harvesting methods are prone to error and are labour-intensive. Recent advances in aerial imagery and Artificial Intelligence (AI) offer innovative solutions to these challenges. An automated yield estimation and fruit harvesting system will require an accurate fruit object detection model. This study explores the application of deep learning models for mango detection on two diverse datasets. This research focuses on evaluating two primary deep learning models: Faster Region-based Convolutional Neural Network (Faster R-CNN) and You Only Look Once (YOLO) variants. Experiments were conducted using datasets from the ACFR mango dataset and a locally collected dataset. The ACFR dataset is acquired from a moving ground vehicle while the local custom dataset is taken from a low flying drone. The Faster R-CNN model was tested with ResNet-50 backbones. YOLOv8 with simple training image augmentation demonstrated superior performance on both datasets, achieving a mean Average Precision with 0.5 Intersection over Union ([email protected]) of 0.959 on the ACFR dataset and 0.756 on the custom local mango image dataset. The YOLOv8 model outperforms Faster R-CNN by a large margin on both datasets. The advantage of simple image augmentation for improving mango detection has also been demonstrated. The YOLOv8 model is found to be able to detect mango fruits effectively in both dark and bright lighting conditions.

Download

Share

COinS