ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Automated Chestnut Grading by Multispectral Vision and Deep Learning

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600081.(doi:10.13031/aim.202600081)
Authors:   Haseeb Amjad, Yuzhen Lu, Yuyuan Tian, Naseeb Singh
Keywords:   Chestnuts, Defect Detection, Deep Learning, Multispectral Imaging

Abstract. Chestnuts are a highly valued specialty crop that requires stringent quality control to meet increasing market demands. Conventional postharvest grading relies heavily on manual inspection, which is labor-intensive, subjective, and difficult to sustain under industrial throughput conditions. Standard color vision systems often face challenges in differentiating defects from the natural dark-textured shell appearance of chestnuts and maintaining stable tracking under high imaging speed and low visual contrast. This study presents a roller-conveyor-based multispectral vision system with advanced deep learning techniques for automated chestnut grading. A five-band multispectral dataset was acquired for chestnuts at conveyor speeds of 10 cm/s and 20 cm/s, comprising red-green-blue (RGB) and two near-infrared (NIR) imagery. A multispectral deep learning framework, ChestnutYOLO, was developed for enhanced chestnut grading. The developed architecture uses a Space-to-Depth (SPD-Conv) stem to retain fine spatial features of small defects and learned 4-channel spectral embedding (R, NIR1, NIR2, DIFF) to improve defect detection. A learnable circular aspect-ratio prior was integrated into the detection head to exploit the characteristic geometry of chestnuts. To account for chestnut rotation during conveyor transport, a temporal confidence voting engine was developed to aggregate defect evidence across multiple views. Experimental results showed that ChestnutYOLO achieved an overall grading accuracy of 94.06% with a full-pipeline throughput of 11.91 frames per second (fps) at the conveyor speed of 10 cm/s and maintained a high grading accuracy of 92.01% at 13.45 fps at the conveyor speed of 20 cm/s. The proposed multispectral framework substantially outperformed RGB-only baselines. Overall, this study demonstrates the potential of multispectral vision with deep learning for non-destructive and automated online chestnut inspection.

(Download PDF)    (Export to EndNotes)