Development and Systematic Evaluation of An Innovative Multispectral Vision-Based Sweetpotato Grading and Sorting System

Sweetpotato quality grading and sorting are still performed predominantly by manual labor at commercial packing facilities. While color imaging-based machine vision systems have been increasingly adopted for automated fruit sorting, multispectral vision that captures information beyond the visible range offers enhanced capacity for fruit defect detection and is well-suited for high-speed automation. Building on our prior work, this study presents the development and systematic evaluation of an innovative multispectral vision-based automated sweetpotato grading and sorting system. The system employs a multispectral vision unit built on a custom-designed roller conveyor that simultaneously captures five-band images, including RGB (red, green, blue) and two near-infrared (NIR1, NIR2) images. A novel three-position rotary pneumatic actuator-based mechanism, fully integrated with the vision unit, sorts the graded sweetpotatoes into three destinations. A deep learning-based pipeline performs real-time segmentation, tracking, and quality assessment of individual sweetpotatoes, with final grades ("Premium", "Good", and "Fair") determined by aggregating multi-view observations captured as sweetpotatoes rotate on the conveyor. Once graded, samples are directed to their respective destinations by the automated sorting mechanism. Experiments were conducted at conveyor speeds of 15, 25, and 35 cm/s to evaluate the vision unit, the sorting mechanism, and the integrated system. The RG+NIR1 band combination paired with the YOLOv11m model consistently delivered the highest grading performance, achieving sample-level accuracies of 97.1%, 94.5%, and 93.6% at the three conveyor speeds, respectively. The sorting mechanism achieved accuracy and repeatability of up to 98.2% and 94.4%, respectively, with a throughput of two sweetpotatoes per second (per lane). The integrated system, using the best-performing grading pipeline, achieved sorting accuracies of 94.1%, 92.8%, and 91.4% at the corresponding speeds. The results demonstrate the effectiveness of multispectral vision and the new sorting mechanisms for high-throughput sweetpotato grading and sorting. The integrated system exhibits strong potential for deployment in commercial packing facilities. The multispectral data and software programs developed in this study will be made publicly accessible.