Abstract

In precision agriculture, accurately estimating the volume of fruits is critical for quality control, yield prediction, and post-harvest management. Current vision-based methods primarily rely on object detection or empirical regression, which inherently lack the necessary geometric accuracy for spherical objects like oranges. This paper presents AGRO-YOLO-V, a hybrid framework that combines pixel-accurate instance segmentation with analytical geometric modeling to achieve high-accuracy single-view volumetric estimation. The system utilizes a modified YOLO-based segmentation backbone to generate precise fruit masks, which are then scaled into a 3D volume using a statistically calibrated area-to-volume scaling law. This hybrid learning-geometry paradigm corrects the systematic errors of bounding-box methods. Tests on a custom orange dataset show that AGRO-YOLO-V achieves a Coefficient of Determination (R2) of 0.94 and stimates volume with a Mean Absolute Percentage Error (MAPE) of 7.39% on a held-out test set. This result represents a 1.66 improvement in error reduction over a rigorously calibrated Bounding-Box baseline (MAPE = 12.24%), confirming the method’s superiority. The suggested framework is a scalable and cost-effective solution for deploying intelligent fruit grading and automated agricultural inspection systems in real time.