Algorithms were developed to process video of maize rows and extract plant features to estimate density and spacing of early growth stage maize plants. The Otsu method was modified to compensate for varied amount of plant segmentation noise presented in different operating conditions. Three features were extracted and used to differentiate between weeds and maize plants: projected plant canopy area, plant length in the image row direction, and perpendicular distance of estimated plant centres from the mean crop row position. Algorithm performance was analysed across three tillage treatments, three growth stages, and three target populations varying from 27 000 to 81 500 plants ha−1. Overall, the algorithm estimated the number of plants in 6·1 m crop row sections with a root mean squared error of 2·1 plants or 8·7% of the mean manual count of 24·1 plants per experimental unit. The mean measurement error was significantly different across tillage treatments, but no evidence of significant differences was found across growth stages and plant populations. The error variance at the vegetative growth stages with the seventh or eight leaf collar visible was significantly higher than that at the growth stages with the third or fourth leaf collar visible. No significant differences were found between mean measured and estimated plant spacing distances.
Available at: http://works.bepress.com/stuart_birrell/61/
This is a manuscript of an article published as Shrestha, D. S., B. L. Steward, and S. J. Birrell. "Video processing for early stage maize plant detection." Biosystems engineering 89, no. 2 (2004): 119-129. DOI: 10.1016/j.biosystemseng.2004.06.007. Posted with permission.