Assessing kernel processing score of harvested corn silage in real-time using image analysis and machine learning

• Published in: Computers and electronics in agriculture 2022-12, Vol.203, p.107415, Article 107415
• Main Authors: Rocha, Eduardo M.C., Drewry, Jessica L., Willett, Rebecca M., Luck, Brian D.
• Format: Article
• Language: English
• Subjects: Corn silage; Image analysis; Kernel processing score; Machine learning
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2022.107415

•Images of corn silage moving at 40 m s−1 were acquired with minimal motion blur.•A camera mounted to a SPFH captured real-time images during harvest.•Various processing rolls gap settings produced statistically different KPS.•Machine learning was trained to detect whole kernels in the material stream.•Machine learning correlated (r(38) = 0.98) with sieve analysis for KPS determination. Whole Plant Corn Silage (WPCS), produced by Self-Propelled Forage Harvesters (SPFH), is among the most common dietary ingredients for dairy cows. Kernel processor settings at harvest can impact WPCS quality and milk production. Having the capability to monitor machine performance and WPCS quality during harvest would allow farmers, nutritionists, and custom harvesters to optimize feed quality as the WPCS enters storage. A real-time quality assessment system was developed - a device to acquire images of chopped and processed silage moving through the spout of a SPFH and an image analysis algorithm to estimate the Kernel Processing Score (KPS) given a set of images. The system represents automation of the quality assessment process and could be incorporated in a real-time machine parameter tuning system. An initial experiment of corn particle detection in high-resolution images of stationary WPCS resulted in good detection performance (precision of 0.7381, recall of 0.9624, and intersection-over-union of 0.7104) but exposed inadequacies in the image labeling procedure and was not able to accurately estimate KPS. An improved camera system was designed and consisted of specialized exposure components to capture images at 121 frames per second. This system was able to acquire pictures silage moving at a rate of 36.9 m s−1. The KPS estimation was done offline and occurred in two steps: whole undamaged kernels were counted in a set of 1,000 images using machine learning; KPS was then estimated using a linear regression model previously obtained with the comparison between mechanical sieving KPS and number of kernels per image (r(38) = 0.967, p