A method for rapid field monitoring of boll opening rate for cotton based on digital imagery

Abstract

Chemical defoliation and ripening are a prerequisite for mechanical harvesting of cotton, and the boll opening rate is a critical determinant of timing and rate of defoliates and ripening agents as well as harvest. Given the low efficiency and poor timeliness of manual determination of boll opening rates, we have developed a rapid method based on digital images. Field images were collected 7 days before and 7, 14, and 21 days after the application of harvest aids, with the boll opening rates (BOR) varying from 25 to 95%. We set four shooting heights, five shooting angles and two shooting directions, and a total of 912 original images (each 5,184×3,456 pixels) were obtained. Actual ground boll opening rates were monitored simultaneously. Each single image was segmented into 500×500 pixels sub-images. The four deep learning networks were used to identify opened and unopened cotton bolls, and YOLOv5 performed best in balancing recognition time and accuracy. To address the issue of boundary boll recognition caused by image segmentation, the original images were segmented into 10 different sizes (100, 200, 300, 400, 500, 600, 700, 800, 900, and 1,000 pixels), and YOLOv5 model was then used to identify bolls in each size of the sub-images. The bounding boxes marking cotton bolls at the same position of two different sizes of sub-images, were combined to obtain new corrected bounding boxes in merged image. Based on the true values of BOR, the best combination of sub-images is 400×400 pixels with 700×700 pixels. This combination was used to examine the recognition results of various shooting parameters, and we found that the optimal shooting height for the digital camera was 20-30 cm above the canopy, with a downward angle of 0-30° (BOR higher than 40%) and 15-30° (BOR lower than 40%) from the horizontal and shooting direction parallel to the planting rows. The method established in this study can enable a less-destructive and rapid detection of BOR in the range of 25 to 95% boll opening rate, with a model R²; value >92% and a relative root mean square error <10%, suggesting its high precision and stability for field application.

Plant phenotyping relevance

綿花のboll opening rateという植物器官・生育状態の形質を、デジタル画像と深層学習で迅速・非破壊に推定する手法を開発し、撮影条件と精度を検証しているため、植物フェノタイピング手法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.