A Fast and Automatic Method for Leaf Vein Network Extraction and Vein Density Measurement Based on Object-Oriented Classification

Abstract

Rapidly determining leaf vein network patterns and vein densities is biologically important and technically challenging. Current methods, however, are limited to vein contour extraction. Further image processing is difficult, and some leaf vein traits of interest therefore cannot be quantified. In this study, we proposed a novel method for the fast and accurate determination of leaf vein network patterns and vein density. Nine tree species with different leaf characteristics and vein types were applied to verify this method. To overcome the image processing difficulties at the microscopic scale, we adopted the remote object-oriented classification method applied comprehensively in the field of remote sensing research. The key to this approach is to determine the universally applicable leaf vein extraction threshold values (scale parameter, shape parameter, compactness parameter, brightness feature, spectral feature and geometric feature). Based on our analysis, the following recommended threshold values were determined: the scale parameter was 250, the shape parameter was 0.7, the compactness parameter was 0.3, the brightness feature value was 230∼280, the spectral feature value was 180∼230, and the geometric feature value was less than 2. With the optimal extraction parameters applied, the extraction precision was above 96.40% on average for the nine species studied. The leaf vein density calculation rate increased by more than 87.3% compared to that of the traditional methods. The results showed that this method is accurate, fast, flexible and complementary to existing technologies. It is an effective tool for the fast extraction of vein networks and the exploration of leaf vein characteristics, particularly for large-scale studies in plant vein physiology.

Teasing out elevational trends in infraspecific Prunus taxa: A vein analysis approach

“…Because much of the discussion on leaf venation analysis has centered around computational methods of network extraction, most quantitative studies on leaf vein parameters have focused on leaves that are more straightforward in terms of vein visibility and leaf size, such as those of Diego and De Bacco (2021), Price et al (2011), and Zheng and Wang (2010). Often samples used are either whole leaves or one‐size‐fits‐all sections for all leaves, often with random sampling from various parts of the leaves (Larese et al, 2014; Zhu et al, 2020). The latter is problematic for comparative studies because venation features change across the surface of a leaf (Nardini et al, 2008; Zwieniecki et al, 2004).…”

Few-leaf Learning: Weed Segmentation in Grasslands

“…Since the manual measurement of leaf veins is a time-consuming and labor-intensive task, it has been of great interest to automate the process. In recent years, automatic leaf vein segmentation has been performed with classic computer vision approaches [8], [9], [10], [11], traditional machine-learning approaches with hand-crafted feature extraction [12], [13], [14] as well as few deep learning approaches [15].…”

Few-Shot Learning Enables Population-Scale Analysis of Leaf Traits in Populus trichocarpa

“…In plant phenotyping, segmentation of individual leaves and their venation has seen sparse attention. In general, existing approaches use (i) experimental methods to chemically clear the leaf lamina and stain the veins to highlight the venation against the background [23,24], (ii) image preprocessing by greyscaling, aggregating specific color channels, or spatial rescaling [23][24][25][26][27], (iii) global filters and morphological operations (e.g., Odd Gabor filters, Hessian matrices, vesselness filters, and region merging) to obtain binary segmentations [23,[25][26][27][28], (iv) ensembles of scales and models to make aggregate predictions [24,28], and (v) require hundreds of manually-annotated training samples to produce accurate segmentation models [24]. However, these commonly encountered steps can bottleneck the scalability and accuracy of image-based plant phenotyping at population scale.…”