Aerial-MangoYOLOv8: An Optimized Mango Detection System for UAV-Based Yield Estimation

Abstract

While vision-based yield estimation systems are widely used for horticulture crops nowadays, accurately predicting yield in unstructured orchards remains a challenge. Automated unmanned aerial vehicle (UAV)-based aerial data collection offers a scalable solution for such environments. This study utilizes UAVs to collect data from a target mango orchard, marking points of interest to ensure periodic video capture along the same path, maximizing coverage. A significant challenge addressed in this work is the detection of small mangoes in UAV-captured videos. While you only look once (YOLO) models, such as the latest YOLOv8, excel in object detection, they struggle with small objects in aerial imagery. To deal with this, the YOLOv8 model is customized with several enhancements, such as integration of the efficient subpixel convolution network to improve the spatial resolution of feature maps and incorporation of attention-based EfficientNet-B7 in the backbone, which improves feature extraction. Furthermore, the bidirectional feature pyramid network in the neck network ensures multiscale feature fusion for detecting objects of varying sizes. It also integrates elliptical-geometric intersection over union to optimize anchor box accuracy and improve object localization. Due to its high efficacy in accurately detecting mangoes in aerial images, the proposed model is named as Aerial-MangoYOLOv8. In addition, an Oriented FAST and Rotated BRIEF feature detector is used for tracking the detected mangoes to predict the yield. Rigorous experimental analysis demonstrates the superiority of the Aerial-MangoYOLOv8 model over other YOLO variants and state-of-the-art detection models.