Machine Learning Based Approach for Detection and Size Estimation for On-field Harvesting and Sorting of Apple

A size estimation methodology is developed with advanced machine vision system for on-site sorting and grading of apples based on size. Accurate estimation of fruit size is essential for precision agriculture, mechanical harvesting, and crop load estimation. Size and maturity are often correlated and can vary among different apple varieties. Hence, it enables selective harvesting, ensuring that only apples that are mature and of right size are plucked by reducing waste while maintaining the general quality of the fruits. The aim of this study was to accurately estimate the size of apples within the tree canopies using stereo vision-based technology. Detection model developed using YOLOv8x-seg and the fruit size is estimated by calculating the dimensions of the detected apple regions. RGB-D camera optimized for generation of depth information of each pixel which is used to calculate distance between camera and each apple. This information further combined with depth data to convert pixel measurements to physical measurements. The conversion of pixel size to physical unit was done using calibration model. The YOLOv8x-seg model achieved mean average precision of 94.35% and an F1 score of 90.71% in detection of apples on test set. The accuracy of apple size estimation on tree canopies is 92.65% and thereby optimizes the need of resources by performing various unit operations simultaneously. The result shows the advancement in agricultural operation (size-based sorting of fruits) during mechanical plucking/harvesting by employing 3d machine vision system.