Mapping and quantifying unique branching structures in lentil (Lens culinaris Medik.)

Abstract

Background: Lentil (Lens culinaris Medik.) is a globally-significant agricultural crop used to feed millions of people. Lentils have been cultivated in the Australian states of Victoria and South Australia for several decades, but efforts are now being made to expand their cultivation into Western Australia and New South Wales. Plant architecture plays a pivotal role in adaptation, leading to improved and stable yields especially in new expansion regions. Image-based high-throughput phenomics technologies provide opportunities for an improved understanding of plant development, architecture, and trait genetics. This paper describes a novel method for mapping and quantifying individual branch structures on immature glasshouse-grown lentil plants grown using a LemnaTec Scanalyser 3D high-throughput phenomics platform, which collected side-view RGB images at regular intervals under controlled photographic conditions throughout the experiment. A queue and distance-based algorithm that analysed morphological skeletons generated from images of lentil plants was developed in Python. This code was incorporated into an image analysis pipeline using open-source software (PlantCV) to measure the number, angle, and length of individual branches on lentil plants.

Results: Branching structures could be accurately identified and quantified in immature plants, which is sufficient for calculating early vigour traits, however the accuracy declined as the plants matured. Absolute accuracy for branch counts was 77.9% for plants at 22 days after sowing (DAS), 57.9% at 29 DAS and 51.9% at 36 DAS. Allowing for an error of ± 1 branch, the associated accuracies for the same time periods were 97.6%, 90.8% and 79.2% respectively. Occlusion in more mature plants made the mapping of branches less accurate, but the information collected could still be useful for trait estimation. For branch length calculations, the amount of variance explained by linear mixed-effects models was 82% for geodesic length and 87% for Euclidean branch lengths. Within these models, both the mean geodesic and Euclidean distance measurements of branches were found to be significantly affected by genotype, DAS and their interaction. Two informative metrices were derived from the calculations of branch angle; 'splay' is a measure of how far a branch angle deviates from being fully upright whilst 'angle-difference' is the difference between the smallest and largest recorded branch angle on each plant. The amount of variance explained by linear mixed-effects models was 38% for splay and 50% for angle difference. These lower R² values are likely due to the inherent difficulties in measuring these parameters, nevertheless both splay and angle difference were found to be significantly affected by cultivar, DAS and their interaction. When 276 diverse lentil genotypes with varying degrees of salt tolerance were grown in a glasshouse-based experiment where a portion were subjected to a salt treatment, the branching algorithm was able to distinguish between salt-treated and untreated lentil lines based on differences in branch counts. Likewise, the mean geodesic and Euclidean distance measurements of branches were both found to be significantly affected by cultivar, DAS and salt treatment. The amount of variance explained by the linear mixed-effects models was 57.8% for geodesic branch length and 46.5% for Euclidean branch length.

Conclusion: The methodology enabled the accurate quantification of the number, angle, and length of individual branches on glasshouse-grown lentil plants. This methodology could be applied to other dicotyledonous species.

Keywords: Image analysis; LemnaTec; Lentil; Morphology; Phenotyping; PlantCV; Python; Structure.

Fig. 1

Lentils growing at “Plant Phenomics Victoria”: A Young lentil plants growing in white 200 mm diameter pots with blue support cages on the LemnaTec Scanalyser 3D phenomics platform; B a typical RGB image of a lentil plant captured by cameras in the LemnaTec Scanalyser 3D phenomics platform; C lentil plants being moved from the glasshouse to the imaging cabinet via conveyors; D mature lentil plants on the LemnaTec Scanalyser 3D phenomics platform at the conclusion of Experiment 1

Fig. 2

Representation of stages in the image analysis pipeline to identify unique branches in lentils: A a cropped section of the original RGB image; B the ‘Y’ channel from a CMYK transformation of the original RGB image, this gives a greyscale representation that is used to help isolate the plant from its background. C A binary image derived from the Y channel, thresholded at grey level 63. D A 9 × 9 pixel dilation applied to the binary image to bind some of the objects; E the ‘M’ channel from CMYK transformation of the original RGB image, this gives a greyscale representation which is used to isolate the cage; F a binary image derived from the M channel, thresholded at grey level 55; G a pixel-wise logical-AND operation applied between the thresholded images of the plant and the cage to identify pixels that overlap the cage and the plant so as to restore the parts of the plant occluded by the cage; H binary image of the plant without the cage; J the ‘K’ channel from a CMYK transformation of the original RGB image that gives a greyscale representation and is used to help isolate any visible potting mix. This is thresholded to create a binary image [not shown]; K the potting mix is subtracted from the plant mask via the application of a pixel-wise logical-AND operation between an inverted potting mix mask and the plant mask; L a 13 × 13 pixel dilation applied to the binary mask; (M) A skeleton derived from the binary shape in L. Note the two disconnected components; N a heatmap showing the distance of each pixel from the top of the pot where greater distance values map to lighter pixel values; P the original RGB image superimposed with unique branch lines derived from the skeleton when all loops are removed, lateral branches clipped and small independent branches restored

Fig. 3

Schematic representation of a simplified binary skeleton that is being processed to identify unique branches which are branches that reach the base (nominally the level of the potting mix). A A skeleton consisting of 23 pixels comprising of one main branch with one left and one right lateral branches. Note that the base of the skeleton is highlighted with a geodesic distance value of 1. Endpoint pixels are displayed in orange, slab pixels are displayed in pink/red and junction pixels are displayed in blue. Each pixel has a geodesic distance value measured from the base point (shown as a number within each pixel). A queue-based algorithm processes each branch sequentially, starting at the endpoint of the longest branch (highlighted with a yellow arrow). The tracking process, which has a current position indicated by a pixel with a blue border, sets each visited pixel to 0 and then progresses down the branch using the 8-connected neighbourhood (dotted orange box) to identify the neighbour with the minimum distance value before moving to that position. B All of the pixels in the main branch have been set to zero. Note that the tracking process terminated at a pixel with the geodesic distance value of 1; therefore, it is a unique branch. The starting condition of the left lateral branch is highlighted. C All of the pixels in the left branch have been set to zero and the tracking process terminates at a pixel with a distance value of 7, which is greater than 1. Note that the starting condition of the right lateral branch is highlighted. D All pixels in the right branch have been set to zero. Note that the right branch terminates at a pixel with a distance value of 4, which is greater than 1

Fig. 4

Representation of a pixel neighbourhood around a pixel at position p. 4-connected pixels have a distance value of 1 (light green) whilst 8-connected pixels have a distance value of √2 (dark green) in accordance with Euclidean geometry

Algorithm 1

Queue based geodesic distance transform

Algorithm 2

Collection of skeleton branches

Fig. 5

Box plots showing the distribution of mean number of branches counted by the lentil branching algorithm in Experiment 1 for each lentil cultivar at 0°, 20°, 40°, 60° and 80° imaging angles which have been pooled. The mean value is indicated with an orange cross. Across the experiment, the image angle at which an image was taken was not found to have a significant effect on branch counts (p = 0.5751)

Fig. 6

Observed mean number of branches per plant over time counted by the branching algorithm in Experiment 1 for lentil plants grouped into classifications based on previously-reported branching phenotypes as shown in the legend. Note that from days 23 to 32 (shaded), the reported classifications accorded with the measurements collected by the algorithm. The phenotype data was collected from field trials over several years conducted by Agriculture Victoria

Fig. 7

Box plots showing the distribution of geodesic lengths of branches counted by the branching algorithm in Experiment 1 for each lentil cultivar. The mean value is indicated with an orange cross. Note that the geodesic distances of branches increase over time for all cultivars and differ between cultivars. The break in the data on days 19 and 20 was due to a mechanical failure which prevented the collection of images on those days

Fig. 8

Box plots showing the distribution of branch splay counted by the branching algorithm in Experiment 1 for each lentil cultivar. The mean value is indicated with an orange cross. This figure shows the variability of branching architecture phenotypes between cultivars

Fig. 9

Box plots showing the distribution of angle difference counted by the lentil branching algorithm in Experiment 1 for each lentil cultivar. The mean value is indicated with an orange cross

Fig. 10

Bar chart showing the distribution of branch counts calculated by the lentil branching algorithm in Experiment 2 for each treatment group on days 14, 18, 25, 28 and 35. Error bars show 95% confidence interval

Fig. 11

A Mean Euclidean branch length, B mean geodesic branch length, C mean branch number and D mean splay (D) calculated by the branching algorithm in Experiment 2 for plants grouped into classifications based on their reported salt tolerance as being “highly tolerant”, “tolerant”, “moderately intolerant”, “intolerant” or “highly intolerant” on days 14, 18, 25, 28 and 35. Error bars show the standard error to a 95% confidence interval