Accurate root length measurement by image analysis

Abstract

Algorithms for estimating root length by image analysis should lead to results that have no systematic error (bias), be insensitive to preferential root orientation, valid across a wide range of sample sizes and adjust for overlap between roots in samples, to reduce the effort needed in spreading out root systems. We propose a new algorithm that forms a compromise between small bias and robustness (insensitivity to variation in sample size and preferential root orientation), and provide a simple way of dealing with root overlap. Image analysis was performed on a Macintosh computer using the public domain NIH Image program. The digital image of the root was processed to get the thinned image (skeleton). The numbers of orthogonally and diagonally connected pairs of pixels (N _o and N _d, respectively) in the skeleton were counted separately and used for length (L) calculation. A new length calculation equation was introduced so that the effect of orientation on length calculation was minimized; L=[N d ²+(N _d+N _o/2)²]^1/2+N _o/2. The maximum error due to orientation of a single line was evaluated for an ideal line, and the analysis revealed that the new equation was less affected by orientation than previous equations. Copper wire and rice (Oryza sativa L.) roots containing both primary and fine secondary root were measured manually and with image analysis. The two methods showed good agreement within 1.5%. The proposed image analysis method yielded length estimates with CV from 0.23 to 0.88%, which was lower than the CVs of the line-intersect method. Moreover, the lengths of overlapping samples were calculated correctly because the image analysis method distinguished an overlapping pixel from a thinned image, while the calculation with the line-intersect method showed underestimation because overlaps were not considered in that method.