Box-counting dimension revisited: presenting an efficient method of minimising quantisation error and an assessment of the self-similarity of structural root systems

• Published in: Frontiers in plant science Vol. 7
• Main Authors: Martin eBouda, Joshua S Caplan, James Edward Saiers
• Format: Journal Article
• Language: English
• Published: Frontiers Media S.A 01.02.2016
• Subjects: fractal dimension; MATLAB code; Numerical Optimisation; Plant root growth; Root system architecture; self-similarity
• DOI: 10.3389/fpls.2016.00149

Fractal dimension (FD), estimated by box-counting, is a metric used to characterise plant anatomical complexity or space-filling characteristic for a variety of purposes. The vast majority of published studies fail to evaluate the assumption of statistical self-similarity, which underpins the validity of the procedure. The box-counting procedure is also subject to error arising from arbitrary grid placement, known as quantisation error (QE), which is strictly positive and varies as a function of scale, making it problematic for the procedure's slope estimation step. Previous studies either ignore QE or employ inefficient brute-force grid translations to reduce it. The goals of this study were to characterise the effect of QE due to translation and rotation on FD estimates, to provide an efficient method of reducing QE, and to evaluate the assumption of statistical self-similarity of coarse root datasets typical of those used in recent trait studies. Coarse root systems of 36 shrubs were digitised in 3D and subjected to box-counts. A pattern search algorithm was used to minimise QE by optimising grid placement and its efficiency was compared to the brute force method. The degree of statistical self-similarity was evaluated using linear regression residuals and local slope estimates.QE due to both grid position and orientation was a significant source of error in FD estimates, but pattern search provided an efficient means of minimising it. Pattern search had higher initial computational cost but converged on lower error values more efficiently than the commonly employed brute force method. Our representations of coarse root system digitisations did not exhibit details over a sufficient range of scales to be considered statistically self-similar and informatively approximated as fractals, suggesting a lack of sufficient ramification of the coarse root systems for reiteration to be thought of as a dominant force in their development. FD estimates did not characterise the scaling of our digitisations well: the scaling exponent was a function of scale. Our findings serve as a caution against applying FD under the assumption of statistical self-similarity without rigorously evaluating it first.