Image processing algorithms for infield single cotton boll counting and yield prediction

• Published in: Computers and electronics in agriculture Vol. 166; p. 104976
• Main Authors: Sun, Shangpeng, Li, Changying, Paterson, Andrew H., Chee, Peng W., Robertson, Jon S.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.11.2019; Elsevier BV
• Subjects: Agricultural production; Algorithms; bolls; Burrs; cameras; color; Color imagery; Cotton; Cotton boll counting; Crop growth; crop management; Crop yield; Digital cameras; Digital imaging; Elongation; Gossypium; growers; growth and development; High throughput phenotyping; Hough transformation; Illumination; Image acquisition; image analysis; Image processing; Image segmentation; Inspection; lighting; Linear Hough Transform; Object detection; Object recognition; Performance evaluation; phenotype; plant breeders; regression analysis; Regression models; Yield estimation; yield forecasting; Image segmentation; Yield estimation; High throughput phenotyping; Linear Hough Transform; Object detection; Cotton boll counting
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2019.104976

•A region growth method was developed to segment cotton bolls in color images.•Disjointed boll regions were merged using line feature and minimum boundary distance.•Bolls overlapping in clusters were split using area and elongation ratio features.•Fiber yield was predicted using cotton boll number. Cotton boll number is an important component of fiber yield, arguably the most important phenotypic trait to plant breeders and growers alike. In addition, boll number provides a better understanding on the physiological and genetic mechanisms of crop growth and development, facilitating timely decisions on crop management to maximize profit. Traditional in-field cotton boll number counting by visual inspection is time consuming and labor-intensive. In this work, we presented novel image processing algorithms for automatic single cotton boll recognition and counting under natural illumination in the field. A digital camera mounted on a robot platform was used to acquire images with a 45° downward angle on three different days before harvest. A double-thresholding with region growth algorithm combining color and spatial features was applied to segment bolls from background, and three geometric-feature-based algorithms were developed to estimate boll number. Line features detected by linear Hough Transform and the minimum boundary distance between two regions were used to merge disjointed regions split by branches and burrs, respectively. The area and the elongation ratio between major and minor axes were used to separate bolls overlapping in clusters. A total of 210 images captured under sunny and cloudy illumination conditions on three days were used to validate the performance of the cotton boll recognition method, with an F1 score of around 0.98; whereas, the best accuracy for boll counting was around 84.6%. At the whole plot level, fifteen plots were used to build a linear regression model between the estimated boll number and the overall fiber yield with a R2 value of 0.53. The performance was evaluated by another ten plots with a mean absolute percentage error of 8.92% and a root mean square error of 99 g. The methodology developed in this study provides a means to estimate cotton boll number from color images under field conditions and would be helpful to predict crop yield and understand genetic mechanisms of crop growth.